Computer Science Dissertation & Thesis Topics
The definitive guide to 130+ computer science dissertation and thesis topics for undergraduate, master’s, and PhD students — spanning AI, machine learning, cybersecurity, NLP, distributed systems, blockchain, quantum computing, HCI, and more. With research gap strategies, methodology frameworks, and writing guidance for every academic level.
💻 Need expert help with your CS dissertation or thesis? Our computer science specialists are ready.
Get Expert Help →What Makes a Strong CS Dissertation Topic — and Why Selection Is the Hardest Part
A computer science dissertation or thesis is a substantial, original research document in which a student identifies an unsolved or inadequately solved problem in a CS sub-field, reviews the existing literature to establish a research gap, applies an appropriate methodology (algorithmic design, empirical evaluation, formal proof, system implementation, or user study), produces and evaluates results, and situates those results within the scholarly conversation. Unlike coursework, a dissertation must make an original contribution to knowledge — at PhD level, that contribution must advance what the field knows; at master’s level, it must demonstrate the ability to conduct independent research; at undergraduate level, it must demonstrate competency in applying research methods to a defined CS problem. The quality of your topic determines the feasibility of that contribution before a line of code is written or a single paper is cited.
The selection of a computer science dissertation topic is where most students either gain a significant advantage or create serious problems for themselves — and the decision is made under conditions that are systematically unfavourable. You are being asked to identify a problem that has not been solved, in a field you are still learning, using methods you may not have fully mastered, within a timeline that is almost always shorter than it should be. The temptation is to choose a topic that feels safe — something well-established, heavily documented, and technically familiar — without realising that well-established topics are well-established precisely because they have been extensively studied, which makes identifying a genuine research gap within them much harder.
The most productive way to approach topic selection in computer science is to invert the usual logic. Rather than starting with a topic and looking for a research question within it, start with a specific problem — a limitation in an existing system, a contradiction in published benchmarks, an underexplored application context, a security vulnerability class that current detection methods do not handle — and work outward from that problem to identify the relevant literature, the appropriate methodology, and the shape of the contribution. This problem-first approach produces naturally researchable topics with built-in novelty, because the problem itself is the research gap. It also produces dissertations that are easier to write, because every section — literature review, methodology, results, discussion — is organised around a problem whose solution the reader already cares about.
The ten domains covered in this guide — artificial intelligence and machine learning, cybersecurity and network security, computer vision and image processing, natural language processing and large language models, distributed systems and cloud computing, blockchain and decentralised systems, quantum computing, human-computer interaction, algorithms and data structures, and bioinformatics — represent the full landscape of contemporary CS research activity. Each section maps the domain’s current frontiers, identifies the research gaps that are most generative for student-level research, and provides specific, researchable topic ideas with the argumentative and methodological angles that distinguish strong dissertations from weak ones. The connections between these domains — how adversarial machine learning connects to cybersecurity, how distributed systems underpin blockchain, how NLP connects to HCI through conversational interfaces — are woven throughout the guide, because the most interesting CS research increasingly crosses sub-field boundaries. For expert support at any stage of your CS dissertation or thesis, the computer science specialists at Smart Academic Writing are available to help.
Dissertation vs. Thesis in Computer Science: Clarifying the Terminology
In the United States, a thesis typically refers to a master’s-level research document (40–100 pages) and a dissertation to a PhD-level document (150–350 pages representing an original contribution to knowledge). In the UK, Australia, and many Commonwealth countries, the terms are often reversed — a dissertation is the undergraduate or master’s document and a thesis is the doctoral work. Throughout this guide, both terms are used to refer to the substantial independent research document required at master’s and PhD level, with context indicating which level applies. Always follow your institution’s terminology conventions precisely.
Artificial Intelligence & Machine Learning Dissertation Topics
Artificial intelligence and machine learning constitute the most active and rapidly evolving sub-field in contemporary computer science, generating an enormous volume of published research — which is both an advantage and a challenge for students selecting dissertation topics. The advantage is that the literature is rich, the tools are mature, and the evaluation frameworks are well-established. The challenge is that the field moves so quickly that a topic that was genuinely novel eighteen months ago may now have multiple published solutions, and the landscape of open problems changes faster than most dissertation timelines. Effective AI/ML dissertation topics in 2026 tend to focus on three recurring themes that are likely to remain generative for several years: reliability and robustness (making AI systems that perform consistently and safely in real-world conditions), efficiency and resource constraints (making AI systems that work in environments without abundant compute), and alignment and interpretability (making AI systems whose behaviour is understandable and controllable). These themes connect the foundational entities of the field — neural networks, gradient descent, loss functions, training data, model architectures — to urgent practical and ethical questions that sit at the intersection of CS and broader societal concerns.
AI & Machine Learning — Dissertation Topics
Robustness, efficiency, alignment, and the frontiers of neural architectures
Federated Learning for Privacy-Preserving Clinical AI
How federated learning frameworks enable collaborative model training across distributed hospital datasets without transferring patient data — evaluating convergence properties, communication efficiency, and model quality under realistic non-IID data distributions in clinical settings.
Research gap: Most federated learning benchmarks use IID splits that do not reflect real clinical heterogeneity; evaluation under realistic skewed distributions remains limited.Adversarial Robustness of Deep Neural Networks in Safety-Critical Systems
Systematic evaluation of adversarial attack types (FGSM, PGD, C&W, AutoAttack) and corresponding defence mechanisms (adversarial training, certified defences, randomised smoothing) in domains including autonomous vehicles, medical imaging, and network intrusion detection.
Research gap: Most robustness evaluations use benchmark datasets that do not reflect the specific distribution shifts and threat models of deployed safety-critical systems.Efficient Neural Architecture Search for Edge Deployment
Developing or evaluating NAS algorithms that simultaneously optimise for accuracy, latency, memory footprint, and energy consumption for deployment on resource-constrained devices including microcontrollers and embedded systems.
Research gap: Existing NAS methods predominantly target GPU inference and rarely account for the specific hardware constraints of ultra-low-power microcontrollers.Explainable AI (XAI) Methods for High-Stakes Decision Systems
Comparative evaluation of post-hoc explanation methods (SHAP, LIME, Grad-CAM, integrated gradients) and inherently interpretable models (decision trees, sparse linear models, attention) across fidelity, stability, and human interpretability dimensions in credit scoring, recidivism prediction, and clinical decision support.
Research gap: Evaluation of explanation quality almost entirely relies on automated fidelity metrics; human-subject studies assessing whether explanations actually support better decisions remain rare.Continual Learning and Catastrophic Forgetting in Production ML Systems
Methods for enabling neural networks to learn from sequential data streams without forgetting previously learned knowledge — evaluating elastic weight consolidation, progressive neural networks, memory replay, and parameter isolation approaches across standard continual learning benchmarks and real-world streaming scenarios.
Research gap: Most continual learning methods are evaluated on computer vision benchmarks; their performance in other sequential domains (tabular data, time-series, NLP) is significantly underexplored.Graph Neural Networks for Drug-Drug Interaction Prediction
Applying GNN architectures (GCN, GAT, MPNN, heterogeneous graph networks) to predict adverse drug-drug interactions from molecular structure and pharmacological knowledge graphs — evaluating against existing polypharmacy datasets with emphasis on rare interactions where data sparsity is a fundamental challenge.
Research gap: Existing DDI prediction models perform well on well-represented drug pairs but systematically fail on novel or rare drugs with limited interaction data — a class imbalance and few-shot learning problem with direct clinical consequences.Reinforcement Learning for Adaptive Network Traffic Management
Training deep RL agents (DQN, PPO, SAC) to make real-time routing and Quality-of-Service decisions in software-defined networks — evaluating against traditional traffic engineering methods under realistic traffic patterns and failure scenarios.
Research gap: RL-based network management is predominantly evaluated in simulation; the sim-to-real transfer problem for network environments with stochastic traffic remains largely unaddressed.Model Watermarking and Intellectual Property Protection for Deep Learning
Developing and evaluating methods for embedding verifiable ownership signatures into neural network weights that survive model fine-tuning, pruning, and distillation attacks — and evaluating the trade-offs between watermark capacity, robustness, and model utility.
Research gap: Most watermarking schemes are evaluated against naive removal attacks; systematic evaluation against adaptive adversaries who know the watermarking scheme is still an open problem.Finding AI/ML Research Gaps: The Benchmark-Then-Benchmark-Again Strategy
The most reliable way to find a genuine AI/ML research gap is to identify a well-regarded benchmark or evaluation suite, run several state-of-the-art methods on it, and ask: in what specific conditions do these methods fail? Fail systematically? Fail in ways that matter for real-world deployment? The answer to those questions is almost always a dissertation topic. The ACM Digital Library at dl.acm.org provides access to the full proceedings of major machine learning and AI conferences (NeurIPS, ICML, ICLR, KDD, AAAI) — reading the “limitations” and “future work” sections of recent papers is the fastest route to identifying problems that are known to exist but have not yet been solved.
Cybersecurity & Network Security Dissertation Topics
Cybersecurity is one of the most professionally consequential and intellectually demanding areas of computer science research, sitting at the intersection of systems programming, cryptography, network engineering, human factors, and policy. The field is driven by an adversarial dynamic — attackers and defenders continuously learn from and adapt to each other — that makes it inherently dynamic and consistently generative of new research problems. The dominant thematic connections in 2026 cybersecurity research link four major entity clusters: threats (malware, ransomware, phishing, supply chain attacks, AI-powered attacks, side-channel attacks), defences (intrusion detection, threat intelligence, secure coding, formal verification, zero-trust architectures), cryptographic primitives (post-quantum cryptography, homomorphic encryption, zero-knowledge proofs, secure multi-party computation), and human factors (security usability, phishing susceptibility, security training effectiveness, organisational security culture). Strong cybersecurity dissertations work at the intersection of at least two of these clusters — connecting a specific threat to a specific defence limitation, or connecting a cryptographic capability to a specific deployment barrier.
Cybersecurity & Network Security — Dissertation Topics
Threat detection, post-quantum cryptography, zero-trust, and adversarial security
Machine Learning-Based Intrusion Detection: Overcoming the Label Scarcity Problem
Developing semi-supervised, self-supervised, or few-shot learning approaches for network intrusion detection that do not require large volumes of labelled attack traffic — a fundamental limitation of current ML-IDS systems given how difficult labelled network security datasets are to produce and maintain.
Research gap: Existing ML-IDS benchmarks use pre-labelled datasets that do not reflect the unlabelled traffic distributions encountered in real SOC environments.Post-Quantum Cryptography Migration: Performance, Compatibility, and Risk Assessment
Evaluating the practical deployment challenges of NIST’s post-quantum cryptographic standards (CRYSTALS-Kyber, CRYSTALS-Dilithium, FALCON, SPHINCS+) in existing TLS/HTTPS infrastructure — measuring performance overheads, certificate size impacts, and backward compatibility issues in realistic network environments.
Research gap: While theoretical properties of PQC algorithms are well-studied, their real-world performance under heterogeneous network conditions and with legacy system constraints is significantly undercharacterised.Zero-Trust Network Architecture: Implementation Frameworks and Empirical Evaluation
Developing implementation blueprints and evaluation frameworks for zero-trust network architectures (ZTNA) in enterprise environments — measuring security improvement, performance overhead, operational complexity, and user experience impact relative to perimeter-based security models.
Research gap: Published ZTNA guidance is largely prescriptive and vendor-driven; independent empirical evaluation of security and usability trade-offs across different implementation approaches is rare.Smart Contract Vulnerability Detection: Static Analysis vs. Fuzzing vs. Formal Verification
Systematic comparison of automated vulnerability detection approaches for Ethereum smart contracts — evaluating Slither (static analysis), Echidna (property-based fuzzing), and Certora/formal methods against a benchmark set of known-vulnerable contracts across vulnerability classes including reentrancy, integer overflow, and access control flaws.
Research gap: Evaluation of smart contract analysis tools typically uses either toy contracts or small benchmark suites; large-scale evaluation on production contracts with known exploitation histories provides a more meaningful and challenging test.LLM-Assisted Malware Generation: Threat Characterisation and Detection Countermeasures
Characterising the extent to which large language models can assist in malware development (script generation, obfuscation, evasion strategy discovery) and evaluating whether existing endpoint detection and response systems can detect LLM-assisted malware using current signatures and heuristics.
Research gap: The cybersecurity research community has primarily studied LLM capabilities in attack contexts qualitatively; systematic empirical evaluation of what current defences can and cannot detect remains limited, partly due to responsible disclosure constraints.Side-Channel Attack Resistance in Modern ARM Processors
Evaluating the effectiveness of hardware and software countermeasures against power analysis, timing, and cache-based side-channel attacks on contemporary ARM Cortex-M series processors used in IoT and embedded security applications — including evaluation of compiler-level countermeasures and constant-time programming techniques.
Research gap: Side-channel evaluation methodology is well-established for smart card and FPGA targets; characterising the same attack surface on modern ARM microcontrollers with architectural features like branch predictors and caches introduces new variables that existing evaluation frameworks do not fully address.Natural Language Processing & Large Language Model Dissertation Topics
Natural language processing is the CS sub-field most visibly transformed by the large language model revolution — a transformation so rapid that the foundational entity relationships of the field have been radically reorganised in just four years. Where NLP research in 2020 was organised around specific task pipelines (named entity recognition, relation extraction, sentiment analysis, machine translation), each with its own model architectures and benchmarks, contemporary NLP is dominated by the question of how pre-trained large language models relate to, subsume, or fail to subsume those specific tasks, and what research questions remain productive given that a single sufficiently large LLM achieves near-state-of-the-art performance on many traditional benchmarks. The most generative NLP dissertation topics in 2026 engage directly with this reorganisation: they identify what LLMs do not do well, characterise the failure modes that systematic evaluation reveals, or address the practical deployment and safety challenges that arise when LLMs are used in real-world applications.
LLM Factual Hallucination: Detection, Mitigation, and the Limits of Retrieval Augmentation
Systematic characterisation of LLM hallucination types (entity, relational, numerical, temporal), evaluation of retrieval-augmented generation (RAG) systems as hallucination mitigation, and the limits of RAG when retrieved context itself is unreliable or contradictory.
Cross-Lingual Transfer for Extremely Low-Resource Languages
Evaluating multilingual model fine-tuning, few-shot prompting, and cross-lingual data augmentation strategies for languages with fewer than 10,000 training examples — connecting transformer architecture to linguistic typology to identify when structural similarity enables effective transfer.
Jailbreaking Resistance: Evaluating LLM Safety Fine-Tuning Under Adversarial Prompting
Systematic red-teaming evaluation of instruction-tuned LLMs using known jailbreak categories — role-playing, many-shot, encoding-based, and multi-turn attacks — to characterise which safety fine-tuning approaches provide genuine safety versus surface-level compliance.
LLM-Based Code Generation: Correctness, Security, and the Automated Testing Problem
Evaluating code generation models (Codex, DeepSeek Coder, StarCoder2, Claude) on functionally complex and security-sensitive programming tasks — not just HumanEval-style competitive programming benchmarks, but real-world tasks involving correct API usage, exception handling, and absence of common vulnerability classes. This connects NLP directly to software engineering and cybersecurity, creating a rich interdisciplinary dissertation scope that reflects the actual deployment context for code-generation AI in software development pipelines. Research gap: HumanEval and MBPP benchmarks focus on algorithmic correctness without testing security properties; evaluation frameworks for code generation that jointly assess functional correctness and security absence are significantly underexplored.
Social Bias in Clinical NLP: Downstream Effects on Healthcare AI Equity
Evaluating how social biases in large clinical language models (BioBERT, ClinicalBERT, MedPaLM) propagate to downstream clinical tasks — medical NER, ICD coding, clinical note summarisation — with particular attention to differential performance across patient demographic groups, connecting NLP bias research to health equity outcomes.
Beyond Perplexity: Evaluation Framework Design for Task-Specific LLM Capability
Developing evaluation frameworks that correlate more reliably with real-world LLM performance on specific tasks than existing benchmarks — addressing the benchmark contamination and saturation problems that make current evaluations unreliable.
Chain-of-Thought Reasoning: When It Works, When It Fails, and Why
Systematic evaluation of chain-of-thought prompting across reasoning task types — mathematical, logical, causal, analogical — characterising the conditions under which CoT improves and degrades accuracy relative to direct prompting.
Multimodal LLMs for Accessibility: Screen Reader Enhancement and Visual Content Description
Evaluating vision-language models for generating high-quality alt-text and screen reader descriptions for complex visual content (charts, diagrams, scientific figures) — connecting NLP research to HCI accessibility requirements.
LLM Quantisation and Pruning: Accuracy-Efficiency Trade-offs Across Deployment Contexts
Evaluating quantisation (GPTQ, AWQ, GGUF) and structured pruning approaches across model families and deployment targets — characterising at what compression ratios specific capability classes degrade, and whether capability degradation is uniform across tasks.
Computer Vision & Image Processing Dissertation Topics
Computer vision — the sub-field concerned with enabling computers to interpret and understand visual information from the world — has been transformed by deep learning architectures (CNNs, transformers, diffusion models) in ways that have simultaneously solved many classical problems and opened new research frontiers. The entity relationships that connect contemporary computer vision research span from sensor inputs (cameras, LiDAR, medical imaging modalities, satellite imagery) through representation learning (convolutional features, vision transformers, CLIP-style contrastive embeddings) to task outputs (classification, detection, segmentation, depth estimation, 3D reconstruction) and deployment contexts (autonomous vehicles, medical diagnosis, surveillance, augmented reality, robotics). Strong vision dissertations identify a specific gap in the chain from input to output in a specific deployment context — not “improve object detection” but “improve small-object detection in aerial imagery for UAV-based infrastructure inspection under challenging lighting conditions.”
Computer Vision & Image Processing — Dissertation Topics
Medical imaging, autonomous systems, generative models, and 3D vision
Foundation Models for Medical Image Segmentation: SAM Adaptation and Domain Gap
Evaluating and adapting Segment Anything Model (SAM) and its medical variants (MedSAM, SAM-Med2D) for clinical segmentation tasks across modalities (CT, MRI, histopathology) — characterising the domain gap between natural and medical image distributions and evaluating fine-tuning strategies that preserve foundation model generality while improving medical domain specificity.
Research gap: SAM’s interaction-based design assumes a human-in-the-loop prompt for every image; adapting it for fully automated clinical workflows without per-image interaction while maintaining accuracy represents a substantial open problem.Deepfake Detection Under Adversarial Generation: Arms Race Evaluation
Systematic evaluation of deepfake detection models when the generative model used to create fakes is specifically optimised to evade detection — characterising whether detection methods based on frequency artefacts, facial inconsistency, and temporal coherence remain robust when the attacker has access to the detector’s architecture.
Research gap: Most deepfake detection benchmarks are created with generative models that are not adversarially optimised against detectors; evaluating the more realistic adversarial setting reveals much larger performance gaps than standard benchmarks suggest.Self-Supervised Learning for Remote Sensing: Downstream Task Transfer Across Sensor Modalities
Pre-training vision transformers on large unlabelled satellite imagery datasets using masked image modelling and contrastive objectives — evaluating transfer performance to downstream tasks (land use classification, building footprint extraction, flood mapping) with limited labelled data, and comparing against supervised baselines and ImageNet-pretrained models.
Research gap: Self-supervised pre-training for remote sensing has been primarily demonstrated on optical imagery; extension to SAR, multispectral, and hyperspectral modalities with their fundamentally different signal characteristics is significantly underexplored.Video Understanding with Efficient Temporal Attention: Long-Range Dependency Modelling
Developing or evaluating efficient attention mechanisms for video transformers that can model long-range temporal dependencies across hundreds of frames without the quadratic attention complexity that makes full self-attention computationally prohibitive for long videos — with evaluation on action recognition, video QA, and surgical procedure understanding benchmarks.
Research gap: Efficient video attention methods are overwhelmingly evaluated on short-clip benchmarks (8–32 frames); their behaviour on long-form videos (minutes to hours) with truly long-range dependencies is not well-characterised.Neural Radiance Fields (NeRF) for Industrial Inspection: Speed, Scale, and Defect Sensitivity
Adapting NeRF architectures for real-time 3D reconstruction and defect detection in industrial inspection contexts — evaluating reconstruction quality, inference speed, and sensitivity to surface defect classes (cracks, corrosion, deformation) against conventional structured light and photogrammetry approaches.
Research gap: NeRF’s long training times and sensitivity to lighting conditions make direct industrial deployment challenging; engineering-focused variants that trade some reconstruction quality for substantially faster training and greater robustness to industrial lighting are underexplored.Distributed Systems, Cloud Computing & Edge Research Topics
Distributed systems research addresses one of the most fundamental and persistent challenges in computer science: how do you design systems in which multiple computers coordinate to achieve goals that no single computer could achieve alone, while managing the inherent unreliability of networks, the impossibility of perfect synchronisation, and the competing demands of consistency, availability, and partition tolerance that the CAP theorem famously constrains? The connections between distributed systems theory and contemporary cloud and edge computing practice create a rich landscape for dissertation research. The core entity relationships span from consistency models (strong consistency, eventual consistency, linearisability, causal consistency) through consensus protocols (Paxos, Raft, Byzantine fault-tolerant variants) and distributed data stores (key-value stores, distributed databases, object storage) to orchestration systems (Kubernetes, container runtimes, service meshes) and edge computing architectures that push processing from centralised data centres toward network-adjacent devices.
Cold Start Latency in Serverless Function Execution: Root Causes, Measurement, and Mitigation
Systematic measurement and characterisation of cold start latency in commercial serverless platforms (AWS Lambda, Google Cloud Functions, Azure Functions) across function languages, memory configurations, and invocation patterns — evaluating mitigation strategies including pre-warming, provisioned concurrency, snapshot-based initialisation, and lightweight runtime approaches. This topic connects serverless computing directly to web application performance, API gateway design, and the economics of cloud computing — a strongly connected set of entities that gives the dissertation natural breadth for literature review while keeping the technical contribution focused on a specific, measurable performance characteristic. Research gap: Cold start latency is widely reported as a problem but is poorly characterised across the heterogeneous hardware and runtime configurations of commercial serverless platforms; most published measurements are outdated or platform-specific.
Hierarchical Inference in Edge-Cloud AI Systems: Latency-Accuracy Trade-offs
Designing and evaluating hierarchical AI inference architectures where lightweight models on edge devices handle common cases and defer to larger cloud models for difficult or uncertain inputs — characterising the accuracy, latency, bandwidth, and privacy implications of different decision boundary policies.
Kubernetes Scheduling Optimisation for ML Workloads
Evaluating and extending Kubernetes scheduling for GPU-intensive ML training jobs — addressing the GPU memory fragmentation, gang scheduling, and workload heterogeneity challenges that make default Kubernetes scheduling suboptimal for ML clusters.
BFT Consensus Protocols for Geo-Distributed Deployments
Evaluating the performance of Byzantine fault-tolerant consensus protocols (HotStuff, PBFT, Tendermint) under realistic geo-distributed latency profiles — quantifying the throughput-latency trade-offs that WAN deployment introduces relative to LAN benchmarks.
Distributed Tracing Overhead and Sampling Strategy Optimisation
Characterising the performance overhead of distributed tracing systems (OpenTelemetry, Jaeger, Zipkin) under high-throughput microservice workloads — evaluating adaptive sampling strategies that preserve anomaly detection capability while minimising observability overhead.
Disaggregated Storage in Cloud-Native Databases: Consistency and Performance
Evaluating disaggregated compute-storage architectures (Amazon Aurora, Google AlloyDB model) against tightly coupled storage approaches — characterising the consistency guarantees, failover behaviour, and performance characteristics under mixed OLTP/OLAP workloads.
Blockchain & Quantum Computing Dissertation Topics
Blockchain and quantum computing represent two of the most distinctive emerging areas for CS dissertation research — both because they attract significant research funding and industrial interest, and because both are at stages of development where the gap between theoretical promise and practical implementation is a rich source of researchable problems. Blockchain research connects distributed systems theory (consensus, Byzantine fault tolerance, consistency models) with cryptography (hash functions, digital signatures, zero-knowledge proofs), economics (mechanism design, token incentives, market microstructure), and software engineering (smart contract security, formal verification). Quantum computing connects theoretical computer science (complexity theory, quantum algorithms) with physics (quantum error correction, decoherence, qubit architectures), cryptography (post-quantum and quantum-resistant algorithms), and systems engineering (quantum hardware calibration, quantum circuit optimisation, hybrid classical-quantum algorithms). The connections between these two domains are themselves a research frontier: post-quantum cryptography for blockchain systems, quantum-resistant smart contract signature schemes, and quantum-enhanced optimisation for consensus are all active areas where a well-scoped dissertation can make a genuine contribution.
Blockchain & Decentralised Systems — Dissertation Topics
Scalability, privacy, DeFi security, and post-quantum resistance
Layer-2 Scaling Solutions: Rollup Comparative Evaluation
Comprehensive empirical comparison of optimistic rollups (Arbitrum, Optimism) and ZK-rollups (StarkNet, zkSync Era, Polygon zkEVM) — measuring throughput, latency, cost, finality time, and security assumptions across realistic DeFi transaction workloads.
Research gap: Existing rollup comparisons are primarily conducted under synthetic benchmarks; evaluation under realistic DeFi workload distributions with their specific transaction complexity patterns is underrepresented.MEV (Maximal Extractable Value) Detection and Mitigation in EVM Chains
Characterising MEV extraction patterns (sandwich attacks, frontrunning, liquidation racing) in production Ethereum transaction data using on-chain analytics — and evaluating proposed mitigation mechanisms (SUAVE, private mempools, commit-reveal schemes) for their effectiveness and economic implications.
Research gap: MEV mitigation proposals are predominantly evaluated through game-theoretic models or small-scale simulations; characterisation against historical on-chain data and evaluation of whether mitigations simply shift extraction to new forms is underexplored.Zero-Knowledge Proof Systems in Blockchain Privacy: Proving Time, Proof Size, and Verifier Complexity
Systematic performance comparison of ZKP systems (zk-SNARK, zk-STARK, PLONK, Bulletproofs) across the proving time, proof size, verifier computation, and trusted setup requirements that determine their practical deployability in privacy-preserving blockchain applications.
Research gap: ZKP system comparisons are often conducted by protocol advocates; neutral empirical evaluation across consistent hardware and task parameters, with particular attention to the proving time bottleneck for mobile provers, is scarce.Quantum Computing — Dissertation Topics
NISQ algorithms, error correction, quantum ML, and hybrid architectures
Variational Quantum Eigensolvers: Circuit Depth vs. Accuracy on NISQ Hardware
Benchmarking VQE circuit depth, noise sensitivity, and optimisation landscape properties on available NISQ hardware (IBM Quantum, IonQ, Rigetti) for small molecular systems — characterising the practical limits of chemical accuracy achievable without error correction and evaluating error mitigation techniques (zero-noise extrapolation, probabilistic error cancellation).
Research gap: Most VQE results are reported on idealised simulators; systematic characterisation of the performance gap between simulation and real hardware across different qubit architectures and connectivity graphs is underexplored.Quantum Machine Learning: Genuine Quantum Advantage or Classical Equivalent?
Systematic evaluation of quantum ML models (quantum kernel methods, quantum neural networks, quantum principal component analysis) against well-tuned classical equivalents on identical datasets — addressing the dequantisation results that suggest many proposed quantum ML advantages can be classically replicated.
Research gap: Many quantum ML advantage claims are made without comparison against modern classical ML baselines using the same computational resources; honest comparison that controls for classical compute budget is rare and technically demanding.Quantum Error Correction Code Comparison on Current Hardware
Implementing and evaluating surface code, steane code, and colour code error correction on available superconducting and trapped-ion quantum processors — characterising which code performs most effectively under the specific noise profiles of current hardware generations.
Research gap: Error correction codes are primarily compared theoretically or on simulators; empirical comparison on physical hardware across different qubit counts and circuit depths provides insights that idealised analysis cannot supply.HCI, UX Research, Algorithms & Data Structure Topics
Human-computer interaction and algorithms represent the most distinctly different methodological traditions in the CS dissertation landscape — HCI relies heavily on qualitative and mixed-methods research rooted in social science traditions, while algorithms research is predominantly formal and mathematical. Both, however, share the commitment to producing results that are rigorous, replicable, and generalisable — and both connect to the broader landscape of CS research in ways that are easily missed if treated as purely self-contained sub-fields. HCI research increasingly intersects with AI (studying how people interact with ML systems, chatbots, and algorithmic decision-makers), security (usable security, privacy-preserving interface design), and accessibility (inclusive design, assistive technology). Algorithms research intersects with AI (efficient gradient computation, attention mechanism complexity, graph algorithm applications in ML), distributed systems (consensus algorithm complexity, streaming algorithms for network analysis), and bioinformatics (sequence alignment algorithms, phylogenetic tree construction). These intersections are where the most interesting dissertation topics often lie.
Human Oversight of Algorithmic Decision Systems: Trust Calibration and Automation Bias
User study investigating how interface design elements (explanation type, confidence display, uncertainty visualisation) affect human decision-makers’ ability to appropriately calibrate trust in AI recommendations — and whether good calibration reduces automation bias without eliminating the efficiency gains of AI assistance.
Conversational AI for Cognitively Accessible Information: LLM-Based Plain Language Generation
Evaluating LLM-based systems for automatically simplifying complex documents (legal text, medical instructions, government forms) to plain language — using both automated readability metrics and user studies with participants with cognitive disabilities or limited literacy.
Passkey Adoption and Usability: Comparing FIDO2 Authentication Across User Populations
Mixed-methods study of passkey authentication adoption — quantifying setup success rates, recovery scenario usability, and cross-device synchronisation challenges across user age groups and technical proficiency levels compared to password managers and SMS OTP.
Approximate Nearest Neighbour Search: Comparative Evaluation for High-Dimensional Vector Databases
Systematic benchmarking of ANN index structures (HNSW, IVF-PQ, ScaNN, DiskANN) across embedding dimensions, dataset sizes, and query throughput requirements relevant to LLM embedding retrieval applications — connecting algorithm design to RAG system performance.
Streaming Algorithms for Network Anomaly Detection Under Memory Constraints
Designing and evaluating sketch-based streaming algorithms (Count-Min Sketch, HyperLogLog variants, reservoir sampling) for real-time detection of traffic anomalies in high-speed networks where storing complete flow records is infeasible — connecting algorithm theory to practical network security.
Scalable Community Detection in Dynamic Social Networks
Developing incremental community detection algorithms that efficiently update community structure in response to edge insertions and deletions without requiring full recomputation — evaluating against static baselines on real-world temporal social network datasets.
HCI Dissertation Methodology: Why User Studies Are Not Optional
HCI dissertations that propose a new interface, interaction technique, or accessibility tool without conducting a user study with actual users are methodologically incomplete — no matter how technically sophisticated the implementation. A user study does not need to be large to be valid: a within-subjects comparison with n=20 participants, appropriate counterbalancing, and a clearly defined task scenario can provide meaningful evidence about usability and user experience. The key methodological decisions — within vs. between subjects, task selection, dependent variable operationalisation, and analysis approach — should be justified by reference to published HCI research methodology (consult CHI and UIST proceedings). For methodology guidance, the IEEE Xplore Digital Library at ieeexplore.ieee.org provides full text of IEEE Transactions on Human-Machine Systems, IEEE Pervasive Computing, and conference proceedings including HPCA, VIS, and VR — essential reading for methodology-level HCI and systems research.
How to Find a Genuine Research Gap: A Practical Framework for CS Students
The research gap is the cornerstone of every dissertation — the specific, identified space between what the existing literature has established and what remains unknown, uncharacterised, or inadequately solved. In computer science, research gaps take several characteristic forms: a method that works well in controlled settings but has not been evaluated in real-world conditions; a technique that was proposed and evaluated in one domain but has not been applied or adapted to a related domain where its effectiveness is unclear; an assumption made in existing work that has not been empirically validated; a claim made in one paper that is contradicted by results from another, creating an inconsistency that needs to be resolved; or a scaling behaviour that is understood at small scale but unknown at the scale relevant for practical deployment. Each of these gap types has a different research strategy, different evidence requirements, and different dissertation structure — but all of them produce genuinely researchable topics that can support an original contribution.
Research Gap Identification Framework for CS Dissertations
Strong and weak formulations across five gap types — with the thinking process behind each
Choosing the Right Methodology: A CS Dissertation Research Design Guide
Methodology in computer science dissertations is not a bureaucratic requirement to be filled by copying a standard template from a previous student’s thesis — it is the set of principled decisions about how you will gather evidence to answer your research question that determines whether your results are credible, reproducible, and meaningful. The choice of methodology follows from the type of research question you are asking: a question about performance requires empirical measurement under controlled conditions; a question about correctness requires formal proof or exhaustive testing; a question about user experience requires a human-subjects study; a question about architectural trade-offs may require a combination of implementation, benchmark evaluation, and formal analysis. Most CS dissertations use a combination of methods, but every method should be chosen because it is the right tool for the specific claim it is being used to support — not because it is what previous students in your department used.
CS Research Methodology Selection Matrix
Which method fits which research question — and what each requires to do it well
A note on reproducibility: computer science is experiencing a reproducibility crisis that parallels, and in some areas exceeds, that in experimental biology and psychology. A significant proportion of published CS papers cannot be reproduced using the information provided — due to unreported hyperparameters, non-public datasets, undisclosed hardware configurations, or training procedures that are described at insufficient detail. Your dissertation should be held to a higher standard: every experiment should be fully described such that a competent CS researcher could reproduce your results without contacting you. This means reporting all hyperparameters, all dataset splits, all hardware specifications, all random seeds, and all preprocessing decisions — ideally with code and data released in a public repository. Not only does this make your dissertation a better scholarly contribution; it also forces the discipline of thinking carefully about what you actually did, which consistently produces better understanding of your own results.
Writing Your CS Dissertation: Structure, Process, and Avoiding the Common Traps
A computer science dissertation has a conventional structure that serves specific functions — not because academic tradition demands it but because each section answers a different question that a competent reader of CS research will ask. The abstract answers: what is the problem, what did you do, what did you find? The introduction answers: why does this problem matter, what is the specific research gap, and what is the paper’s contribution? The related work / literature review answers: what has been done before, and how does your work relate to, extend, or improve upon it? The methodology / experimental setup answers: exactly how did you generate your results? The results answers: what did you find? The discussion answers: what do the results mean, what are their limitations, and what do they imply for future work? The conclusion answers: what should the reader take away from this work?
Write the Introduction Last (or at Least Twice)
The introduction claims to introduce what the dissertation will do — but you cannot fully know what you will do until you have done it. Write a draft introduction early to clarify your thinking, then rewrite it after your results are in. The second version is almost always dramatically better because it can be specific about contribution rather than promissory about intent. The most common introduction failure in CS dissertations is promising too much too vaguely: “This dissertation will improve machine learning” is not a contribution statement. “This dissertation proposes and evaluates a novel attention mechanism that reduces transformer inference latency by 34% on the BERT-base architecture with less than 1% accuracy degradation on GLUE” is a contribution statement.
Literature Review: Synthesise, Don’t Annotate
A literature review is not a chronological list of papers with summaries. It is a thematic argument about what the existing research has established, where it is inconsistent or limited, and how your work fills a specific gap in it. Organise your literature review around themes and entities — the major approaches to your problem, the key debates, the evaluation methodologies used, the datasets employed — not around individual papers. Each paragraph should make a claim about the state of knowledge in your area and support it by reference to multiple papers, not one-paragraph-per-paper summaries that read like annotated bibliographies.
Experimental Setup: Write It So Anyone Can Reproduce
The methodology and experimental setup section is the most important section for CS research integrity. Every decision that affects your results must be documented: hardware (CPU, GPU, RAM, storage), software (framework versions, library versions, operating system), dataset (source, version, split proportions, preprocessing steps), hyperparameters (all of them, not just the ones you tuned), evaluation metrics (definition, implementation, aggregation method), and statistical analysis approach (significance tests, confidence intervals, correction for multiple comparisons). If you cannot reproduce your own results from your methods section, the section is inadequate.
Results: Report What You Found, Not What You Hoped to Find
Results sections in CS dissertations should present findings honestly — including null results, negative results, and results that do not support your initial hypothesis. Negative results are not dissertation failures; they are contributions to knowledge. If your proposed method does not outperform the baseline, characterise precisely why it does not, under what conditions it does or does not offer advantages, and what that tells us about the problem. A dissertation that reports honest negative results with good characterisation of why they occurred is more valuable scientifically than one that cherry-picks conditions under which a positive result can be manufactured.
Discussion: Interpret, Qualify, and Situate
The discussion is where the intellectual work of the dissertation becomes visible — where you interpret your results in light of existing knowledge, acknowledge the limitations of your evaluation, and draw conclusions that are appropriately qualified by those limitations. Avoid overclaiming: “Our results demonstrate that X is always better than Y” is almost never justified by a dissertation’s evaluation; “Our results suggest that X outperforms Y under conditions A, B, and C with the caveat that we have not evaluated D” is honest and specific. The discussion should also identify threats to validity — confounds, generalisability limitations, dataset artefacts, and evaluation protocol choices that might have influenced your results.
Citations and Technical Writing: Precision Is Professionalism
CS dissertations typically use numeric citation styles (IEEE or ACM) or author-date (APA for some programmes). Beyond citation format, technical writing precision matters enormously: distinguish between “accuracy” and “precision” (different metrics), “latency” and “throughput” (different performance characteristics), “training data” and “test data” (different evaluation roles), and “model” and “system” (different levels of abstraction). Imprecise technical language signals to examiners that the student does not fully understand what they are measuring or building — and it makes claims harder to evaluate. If you need support ensuring technical writing accuracy and citation correctness, our CS dissertation specialists can help.
Finding CS Dissertation Sources: Databases, Venues, and Search Strategies
The quality of a CS dissertation is inseparable from the quality of its literature — and the quality of its literature depends on knowing where to look, how to search, and how to distinguish primary conference and journal publications from secondary sources, preprints, and blog posts. Computer science is unusual among academic disciplines in that conference proceedings are the primary venue for cutting-edge research — the best CS research appears at conferences (NeurIPS, CVPR, ICML, CCS, SOSP, OSDI, CHI, PLDI, FOCS) before or instead of in journals, and a literature review that covers only journals will systematically miss the field’s most significant recent contributions.
| Resource | Primary Use | Coverage | Access Notes |
|---|---|---|---|
| ACM Digital Library (dl.acm.org) | Full text of ACM conference proceedings and journals — the essential source for systems, HCI, programming languages, and networking | ACM conferences (SIGCOMM, SOSP, CHI, CCS, PLDI, STOC, SIGGRAPH) and ACM journals | Most universities have institutional access; some content freely available via Open Access |
| IEEE Xplore (ieeexplore.ieee.org) | Full text of IEEE conference proceedings and transactions — the essential source for hardware, networking, computer vision, and systems | IEEE conferences (CVPR, ICCV, ICSE, HPCA, VIS, IPDPS) and IEEE Transactions journals | Institutional access required for most content; some open access papers available |
| arXiv (arxiv.org) | Preprint repository for CS, AI/ML, and computational research — often the fastest route to recent results before formal publication | All CS sub-fields; AI/ML papers typically appear on arXiv months before formal conference publication | Fully free; note that preprints are not peer-reviewed — check whether a preprint has a corresponding peer-reviewed publication before citing it as established work |
| Semantic Scholar (semanticscholar.org) | AI-powered literature search with citation network analysis — particularly useful for finding papers that cite a key paper (forward citation) or identifying highly-cited papers in a sub-field | Broad CS and AI coverage with citation graph analytics | Free; integrates with arXiv, ACM, and IEEE sources |
| Google Scholar (scholar.google.com) | Broadest coverage for initial literature discovery — useful for finding all formats of CS publications but requires verification of source quality | All venues including conference papers, journals, preprints, theses, and technical reports | Free; use it for discovery but verify quality by finding the paper’s primary venue |
| Papers With Code (paperswithcode.com) | Finding state-of-the-art results with associated code repositories for ML and AI benchmarks — essential for understanding the current performance frontier and identifying evaluation standards in your area | ML/AI methods with benchmark leaderboards and code repositories | Free; enables rapid survey of performance across methods on standard benchmarks |
Literature Search Strategy: The Snowball Method for CS
The most reliable CS literature search strategy combines forward and backward citation snowballing from a small set of seed papers. Step 1: Identify 3–5 highly-cited papers that are central to your research area using Google Scholar or Semantic Scholar keyword searches. Step 2: Read each paper’s related work section to identify the 10–15 papers they consider most important — these are your backward citations. Step 3: Use Semantic Scholar or Google Scholar to find papers that have cited your seed papers since their publication — these are your forward citations and are especially valuable for identifying recent work. Step 4: Repeat the process with your most important newly discovered papers. This produces a literature set that is both comprehensive and specifically relevant to your problem, rather than the superficial keyword-matching set that database search alone produces.
Essential Vocabulary for CS Dissertation Writing and Searching
Key Terms Across CS Dissertation Domains — For Literature Search and Technical Writing
10 CS Dissertation Mistakes That Cost You Marks — and How to Fix Each
| # | ❌ The Mistake | Why It Weakens the Dissertation | ✓ The Fix |
|---|---|---|---|
| 1 | Choosing a topic that is a description, not a research question | “Exploring machine learning applications in healthcare” is a topic description. It could describe a hundred different research projects, most of which would not constitute a dissertation contribution. Without a specific, researchable question — one that has a non-obvious answer that requires evidence to establish — there is no dissertation to write, only a survey to compile. | Reframe every topic description as a specific question: “Under what conditions does federated learning achieve comparable model quality to centralised training for clinical NLP tasks, and what are the privacy-utility trade-offs?” Now you have a question with an answer that requires empirical investigation. |
| 2 | Literature review that lists papers without synthesising them | An annotated bibliography is not a literature review. A literature review must argue something about the state of existing knowledge — identifying what the field agrees on, where it is divided, what methods dominate and why, and most critically, what gap your work fills. Without this argumentative structure, the literature review does not justify the dissertation’s existence. | Reorganise your literature review around thematic claims: “Existing ML-IDS systems achieve high accuracy on pre-labelled benchmark datasets but perform substantially worse in deployment settings where labelled attack data is scarce [citations] — a gap that no existing approach directly addresses” is a claim your literature review should make and support, not a comment you make after summarising each paper. |
| 3 | Evaluation against outdated or inappropriately chosen baselines | A new method that outperforms baselines from 2018 in a field where 2023 models exist has not demonstrated meaningful improvement. Examiners and reviewers will immediately identify which baselines were available when your work was conducted, and failing to include strong recent baselines signals either ignorance of the literature or deliberate selection of weak comparison points. | Before finalising your evaluation design, verify that your baselines represent the current state-of-the-art in your specific task and evaluation setting. Papers With Code (paperswithcode.com) provides current benchmark leaderboards for most ML tasks — use them to identify what you must beat to claim state-of-the-art performance. |
| 4 | No ablation study for system or model contributions | If you propose a system with three components and show it outperforms a baseline, the reader cannot determine which component is responsible for the improvement — whether all three are necessary, whether one does all the work, or whether one hurts and the others compensate. Without ablation, the contribution is insufficiently characterised to be actionable for other researchers. | For every component of your proposed system or method, include an ablation that removes or simplifies that component and measures the resulting performance change. The ablation study is not a supplementary analysis — it is the mechanism by which you claim that your proposed components actually cause the observed improvement. |
| 5 | Claiming statistical significance without appropriate tests | Reporting that “method A achieved 94.2% and method B achieved 93.8%” without a statistical significance test does not establish that the difference is meaningful rather than due to random variation in evaluation. This is especially problematic in ML, where model performance varies across random seeds, dataset splits, and initialisation choices. | Report all numerical results with confidence intervals or standard deviations over multiple runs. Use appropriate statistical tests (paired t-test for within-subject comparisons, Wilcoxon signed-rank for non-parametric comparisons) and report p-values. Apply corrections for multiple comparisons (Bonferroni, Holm) when comparing across many conditions. Mean performance over multiple random seeds is a minimum standard for any ML evaluation. |
| 6 | Conflating a tool or implementation with a research contribution | “I implemented X using Y and Z libraries” is an engineering contribution, not a research contribution. A research contribution requires you to have learned something from the implementation that was not previously known — a performance characterisation, a design trade-off analysis, a security evaluation, a usability finding, or a correctness proof. The implementation is the method; it is not itself the result. | Ask: “What do we know after this dissertation that we did not know before, and what evidence supports it?” The answer should be something specific and non-obvious — a finding about behaviour, performance, trade-offs, or design principles. If the answer is “that it is possible to implement X,” that is not a dissertation contribution unless the impossibility of X was previously assumed. |
| 7 | Dataset without ethics approval or data governance analysis | CS dissertations that collect human data (user studies, surveys, crowdsourced annotation, social media scraping) without ethics board approval are not submittable at most universities and may not be publishable. CS dissertations that use existing datasets without understanding their license, provenance, and known biases may produce results that misrepresent real-world performance. | Get ethics approval before any human subjects data collection — this process takes weeks and must be planned before your research begins. For every dataset you use, document its source, license, size, collection methodology, known biases, and any published critiques of its representativeness. This documentation belongs in your methodology section. |
| 8 | Security claims without a threat model | Security dissertations that claim a system “is secure” without specifying a precise threat model — what the attacker can and cannot do, what assets are being protected, what the security goal is — are not making a meaningful scientific claim. Security properties are always relative to a specific threat model; a system that is secure against a passive eavesdropper may be completely insecure against an active man-in-the-middle attacker. | Every security claim in your dissertation must be preceded by an explicit threat model statement: who is the attacker, what capabilities do they have (black-box vs. white-box, adaptive vs. non-adaptive), what is the attacker’s goal, and what is the security property you are claiming to achieve? Only then can you evaluate whether your system achieves that property against that attacker. |
| 9 | Discussion section that just repeats the results section | “The results show that our method achieved 94.2% accuracy while the baseline achieved 91.5%” is a result, not a discussion. A discussion should interpret what the results mean in the context of the research question, explain why the results came out the way they did, acknowledge their limitations, and situate them in the broader scholarly conversation. Repeating results without interpretation is a sign that the student has done the experiment but not the thinking. | For every major result, ask three questions in your discussion: (1) Why did this happen? What mechanism explains the performance difference? (2) When does this not hold? Under what conditions might the result be reversed? (3) What does this mean for the field? What implication does this finding have for related work, for practical deployment, or for future research directions? |
| 10 | Ignoring reproducibility in code and experiment reporting | A dissertation in which the experiments cannot be reproduced has contributed nothing reproducible to scientific knowledge — the findings exist only on paper and cannot be built upon. CS research has a documented reproducibility crisis, and dissertations that contribute to it are justifiably criticised by examiners and future researchers alike. | Release your code and data publicly (GitHub, Zenodo, Hugging Face Hub) and include the link in your dissertation. At minimum, report every hyperparameter, every preprocessing decision, every hardware specification, and every random seed that affects your results. A reproducibility appendix that lists everything needed to reproduce each result figure is a standard of academic quality that your dissertation should aspire to. |
Pre-Submission CS Dissertation Checklist
✓ Research Quality
- Research question is specific, researchable, and not a topic description
- Literature review synthesises — does not just annotate individual papers
- Baselines represent current state-of-the-art at time of evaluation
- Ablation study identifies the contribution of each system component
- Statistical significance testing applied with appropriate tests and corrections
- Threat model explicitly stated for any security claim
- Ethics approval obtained for all human subjects research
- Negative results reported honestly with characterisation of why they occurred
✓ Writing & Reproducibility
- Contribution statement is specific and concrete — not promissory
- All hyperparameters, datasets, and hardware documented in methodology
- Code and data released publicly with link in dissertation
- Technical terminology used precisely and consistently throughout
- Discussion interprets results — does not just repeat them
- Limitations section acknowledges specific threats to validity
- Citation style (IEEE, ACM, or APA) applied consistently
- All figures have axis labels, units, and are readable in greyscale
FAQs: CS Dissertation & Thesis Topics Answered
Choosing Well, Researching Rigorously, and Writing Honestly
A computer science dissertation is not a demonstration that you can implement something — it is a demonstration that you can think like a researcher. That means identifying a problem that matters, understanding what the existing literature has and has not established about that problem, designing an investigation that will produce meaningful evidence, conducting that investigation with methodological rigour, and communicating your findings with the precision and honesty that scientific writing demands. Each of those steps is learnable, and this guide has provided frameworks for the most difficult of them: finding your research gap, scoping your topic, designing your methodology, and avoiding the mistakes that most reliably weaken CS dissertations.
The semantic connections between CS sub-fields that run through this guide are not incidental — they reflect the actual structure of contemporary CS research, in which the most interesting problems consistently sit at boundaries: between AI and security (adversarial ML, privacy-preserving learning), between distributed systems and cryptography (blockchain consensus, secure multi-party computation), between NLP and HCI (conversational interfaces, accessibility tools), between algorithms and real-world systems (approximate query processing, streaming network analysis). If your dissertation topic sits at one of these intersections, it is almost certainly more interesting, more publishable, and more impactful than a topic that stays entirely within a single sub-field — because the intersection is where the field’s unsolved problems tend to concentrate.
If you need expert support at any stage of your CS dissertation — from initial topic scoping to final pre-submission polishing — the computer science specialists at Smart Academic Writing are ready to help. Explore our computer science assignment help, our dissertation and thesis writing service, our research paper writing services, our data analysis and statistics support, and our editing and proofreading service. Learn how we work, review our transparent pricing, and contact us to get started on your CS dissertation today.
