Computer Science Research Topics

What Makes a Strong Computer Science Research Topic?

Computer science as a discipline spans an extraordinary range — from the most abstract mathematical investigations into computability, complexity, and formal logic, through the engineering science of systems, networks, and hardware, to the deeply human-facing questions of usability, accessibility, and the social impact of technology. Choosing a research topic means first understanding which part of this spectrum you inhabit, because the research methods, publication venues, evaluation criteria, and writing conventions differ substantially between, say, a paper on quantum algorithm design and a paper on the effects of dark patterns in user interface design.

According to the ACM Digital Library — the world’s largest collection of CS research literature — computer science publications span more than fifty distinct sub-fields, with over 500,000 new articles and conference papers published annually. Navigating this landscape as a student researcher requires knowing not just what you want to study, but where in the literature your work will be situated, what methodological standards your sub-field expects, and which conferences and journals your target community reads.

Artificial intelligence and machine learning are the fastest-moving areas of computer science research — with publication velocity so high that a survey paper published six months ago may already be significantly outdated. The sub-field spans foundational theory (learning theory, optimisation, generalisation bounds), systems (distributed training, efficient inference, hardware-software co-design), applications (computer vision, natural language processing, bioinformatics, robotics), and the increasingly urgent questions of safety, alignment, fairness, and interpretability. The breadth of open problems makes AI/ML one of the most accessible research domains for students at every level — but it also makes the choice of a specific, feasible research question more critical than in narrower domains.

• 01
  Interpretability Methods for Large Language Models: Attention Attribution vs. Probing Classifiers Comparing mechanistic interpretability approaches to understand what information LLMs encode in their internal representations and whether attention weights reliably indicate factual knowledge location. Research Question: Do attention attribution scores in transformer-based LLMs correlate reliably with probing classifier accuracy for specific factual knowledge types, and how does this relationship change across model scales?
  PhD / MS
• 02
  Federated Learning Under Non-IID Data: Personalisation Strategies Investigating how data heterogeneity across federated clients degrades global model accuracy and evaluating personalisation mechanisms — including local fine-tuning, meta-learning, and clustered federation — against this degradation. Research Question: Which personalisation strategy most effectively mitigates accuracy loss due to statistical heterogeneity in federated learning, and what dataset characteristics determine the optimal strategy choice?
  Graduate
• 03
  Adversarial Robustness of Vision-Language Models Investigating how multimodal models like CLIP and DALL-E respond to adversarial perturbations applied to the visual modality, and whether text-based grounding provides any robustness benefit. Research Question: Are vision-language models more robust to visual adversarial perturbations than unimodal vision models when the text modality is available, and what is the attack surface of the cross-modal fusion mechanism?
  Graduate
• 04
  Efficient Fine-Tuning of Large Language Models: LoRA vs. Adapter Layers Comparing parameter-efficient fine-tuning methods for LLMs across different task types, model sizes, and data regimes to identify the conditions under which each method achieves the best accuracy-efficiency trade-off. Research Question: Under what combinations of task type, available fine-tuning data size, and base model scale does Low-Rank Adaptation (LoRA) outperform adapter-based fine-tuning on downstream benchmark performance?
  Undergrad
• 05
  Hallucination Detection and Mitigation in Generative AI Systems Developing and evaluating automated methods for detecting factually incorrect outputs from large language models, and assessing whether retrieval-augmented generation (RAG) reduces hallucination rates across domain types. Research Question: Does retrieval-augmented generation significantly reduce hallucination rates in LLMs on domain-specific factual benchmarks, and does the retrieval corpus quality affect the magnitude of this reduction?
  Graduate
• 06
  Algorithmic Bias in Facial Recognition Systems Across Demographic Groups Measuring and characterising performance disparities in commercial and open-source facial recognition systems across race, gender, and age demographics, and evaluating debiasing techniques. Research Question: To what extent do current facial recognition systems exhibit intersectional bias (across combinations of race, gender, and age), and does dataset rebalancing or adversarial debiasing more effectively reduce this intersectional disparity?
  Undergrad
• 07
  Reinforcement Learning from Human Feedback (RLHF): Reward Model Alignment Examining how reward models trained on human preference data encode and generalise alignment signals, and whether reward hacking emerges systematically across policy optimisation steps. Research Question: Does reward hacking — where policy optimisation maximises learned reward at the expense of true human preference — emerge consistently across RLHF implementations, and what training dynamics predict its onset?
  PhD
• 08
  Graph Neural Networks for Drug-Target Interaction Prediction Applying and benchmarking graph-based deep learning architectures for predicting novel drug-target binding affinities using molecular graph representations, comparing against fingerprint-based and sequence-based baselines. Research Question: Do message-passing graph neural networks trained on molecular interaction graphs outperform ECFP fingerprint-based models for drug-target interaction prediction on the BindingDB benchmark?
  Graduate
• 09
  Catastrophic Forgetting in Continual Learning Systems Investigating the mechanisms of catastrophic forgetting in neural networks during sequential task learning and evaluating the effectiveness of regularisation-based, memory-replay, and architecture-based approaches to continual learning. Research Question: Do elastic weight consolidation (EWC) and gradient episodic memory (GEM) mitigate catastrophic forgetting equivalently across task sequence length, or does performance diverge as the number of learned tasks increases?
  Graduate
• 10
  Zero-Shot and Few-Shot Learning with Foundation Models Characterising the conditions under which foundation models generalise to unseen tasks without task-specific training data, and measuring how prompt engineering affects zero-shot performance across task types. Research Question: Does chain-of-thought prompting improve few-shot performance of large language models consistently across reasoning task types, or is the benefit specific to certain categories of multi-step reasoning?
  Undergrad
• 11
  Neural Architecture Search for Edge-Constrained Deployment Developing automated neural architecture search methods that optimise for accuracy under strict memory and latency constraints for deployment on resource-limited edge devices. Research Question: Can multi-objective NAS with explicit hardware-aware latency constraints on target edge processors achieve pareto-optimal accuracy-efficiency trade-offs superior to post-training quantisation of general-purpose architectures?
  PhD
• 12
  Synthetic Data Generation for Privacy-Preserving Machine Learning Evaluating GAN-based and diffusion-model-based synthetic data generation pipelines for healthcare and financial datasets, measuring the fidelity-privacy trade-off using differential privacy metrics. Research Question: Does synthetic training data generated by diffusion models with differential privacy guarantees preserve sufficient statistical fidelity to enable clinical prediction models with performance equivalent to models trained on real patient data?
  Graduate
• 13
  Multi-Agent Reinforcement Learning for Cooperative Tasks Designing and benchmarking MARL algorithms for cooperative multi-agent settings such as autonomous vehicle coordination, warehouse robotics, and network resource allocation under partial observability. Research Question: Does centralised training with decentralised execution (CTDE) in multi-agent reinforcement learning produce stable cooperative policies in environments with highly dynamic agent populations?
  PhD
• 14
  Fairness Metrics and Trade-offs in Predictive Policing Algorithms Analysing the mathematical incompatibility of competing fairness metrics (demographic parity, equalised odds, calibration) when applied to recidivism prediction and predictive policing systems. Research Question: Are demographic parity and equalised odds simultaneously achievable in recidivism prediction models when base rates differ across demographic groups, and how do different fairness interventions affect overall accuracy?
  Undergrad
• 15
  Transfer Learning Efficiency in Low-Resource NLP for African Languages Investigating the effectiveness of multilingual pre-trained language models for NLP tasks in low-resource African languages, and evaluating data augmentation strategies for bridging the resource gap. Research Question: Does cross-lingual transfer from high-resource Bantu language models to low-resource Bantu variants outperform training from scratch given equivalent target-language data budgets?
  Graduate
• 16
  Explainability of Clinical Decision Support Systems Evaluating SHAP, LIME, and counterfactual explanation methods for clinical AI systems, measuring both technical fidelity and clinician comprehension and trust of the explanations produced. Research Question: Do SHAP-based explanations for clinical AI predictions meaningfully increase clinician agreement with correct AI recommendations, and does explanation quality moderate this effect?
  Graduate

Cybersecurity research is defined by its adversarial nature — unlike most engineering disciplines, the “environment” actively tries to subvert your system. This adversarial dynamic creates a rich landscape of open research problems at every level: from the theoretical (provably secure cryptographic protocols, formal verification of security properties) to the applied (malware detection, network intrusion response, web application security) to the socio-technical (human factors in security, usability of privacy tools, policy and governance). Cybersecurity research is published primarily through venues including IEEE S&P (Oakland), ACM CCS, USENIX Security, and NDSS — all of which have published proceedings accessible through the IEEE Xplore Digital Library.

• 17
  AI-Driven Intrusion Detection: Reducing False Positives in Network Security Developing and evaluating machine learning-based network intrusion detection systems (NIDS) that maintain high detection rates while reducing operationally disruptive false positive rates on contemporary network traffic datasets. Research Question: Can anomaly-based NIDS using ensemble learning achieve a false positive rate below 0.1% on the CIC-IDS-2018 benchmark while maintaining detection rates above 95% for zero-day attack variants?
  Graduate
• 18
  Ransomware Behaviour Analysis and Early Detection via System Call Monitoring Using dynamic analysis of system call sequences to identify behavioural signatures of ransomware encryption phases before significant file damage occurs, evaluated on real-world ransomware families. Research Question: Can system call sequence analysis using LSTM-based classifiers detect ransomware encryption behaviour within the first 10 seconds of execution while maintaining a false negative rate below 5% across diverse ransomware families?
  Graduate
• 19
  Post-Quantum Cryptography: NIST Finalist Algorithm Performance Benchmarking Implementing and benchmarking NIST-selected post-quantum cryptographic algorithms (CRYSTALS-Kyber, CRYSTALS-Dilithium, SPHINCS+) on constrained IoT hardware platforms against current RSA and ECC implementations. Research Question: What are the key generation, encapsulation, and decapsulation latency penalties of CRYSTALS-Kyber compared to RSA-2048 on ARM Cortex-M4 microcontrollers, and are they operationally acceptable for IoT authentication scenarios?
  PhD / MS
• 20
  Software Supply Chain Security: Detecting Malicious Packages in Open-Source Ecosystems Analysing the characteristics of known malicious npm and PyPI packages to develop static and behavioural detection heuristics for identifying supply chain attacks before package installation. Research Question: What combination of static code features and package metadata characteristics most reliably distinguishes malicious npm packages from benign ones with minimal false positives on a labelled dataset of known supply chain attacks?
  Undergrad
• 21
  Differential Privacy in Machine Learning: Utility-Privacy Trade-off Analysis Empirically characterising how different privacy budget settings (ε) in differentially private SGD affect model accuracy across datasets of different sensitivity and size, and evaluating membership inference attack success rates. Research Question: At what differential privacy budget (ε) does DP-SGD training prevent membership inference attacks while maintaining classification accuracy within 5% of non-private baseline performance on medical imaging datasets?
  Graduate
• 22
  Phishing Detection Using NLP and URL Feature Analysis Building and evaluating machine learning models combining URL lexical features, HTML content analysis, and domain reputation signals for phishing website detection, including adversarially robust evaluation. Research Question: Do phishing detection models using BERT-based URL encoders trained on the PhishTank dataset significantly outperform traditional feature-engineering approaches when evaluated against adversarially obfuscated phishing URLs?
  Undergrad
• 23
  Side-Channel Attacks on Deep Learning Inference Accelerators Investigating power and electromagnetic side-channel leakage from neural network inference on FPGA and ASIC accelerators, and evaluating countermeasures including masking and noise injection. Research Question: Can power side-channel analysis extract model architecture parameters from black-box FPGA-based DNN inference accelerators, and does differential power analysis resistance transfer from classical cryptographic hardware to neural network settings?
  PhD
• 24
  Zero-Trust Architecture Implementation in Enterprise Networks Evaluating the operational and security impact of transitioning from perimeter-based to zero-trust network architectures in enterprise settings, including identity verification overhead and lateral movement prevention. Research Question: Does zero-trust microsegmentation significantly reduce the blast radius of successful credential-compromise attacks in enterprise networks, and what are the measurable latency costs of continuous authentication verification?
  Graduate
• 25
  Vulnerability Detection in Smart Contracts Using Static Analysis Developing and benchmarking automated static analysis tools for detecting common Solidity smart contract vulnerabilities (reentrancy, integer overflow, access control flaws) against known-vulnerable contract datasets. Research Question: Do existing static analysis tools for Solidity smart contracts achieve acceptable recall on the SWC Registry vulnerability taxonomy, and which vulnerability categories remain systematically under-detected?
  Undergrad
• 26
  IoT Botnet Detection via Network Traffic Behavioural Fingerprinting Building network-level detection models for IoT botnet activity that do not require device access, using only traffic metadata and flow statistics observable at the network perimeter. Research Question: Can flow-level traffic features derived from IoT network captures detect C2 communication patterns of Mirai-variant botnets with sufficient accuracy for operational deployment in residential ISP settings?
  Graduate
• 27
  Usable Security: Password Manager Adoption Barriers and Nudge Interventions Studying the psychological and usability barriers preventing users from adopting password managers despite security awareness, and evaluating the effectiveness of interface nudges and security messaging. Research Question: Which categories of usability friction — setup complexity, cross-device synchronisation, and autofill reliability — most strongly predict password manager non-adoption among security-aware users?
  Undergrad
• 28
  Deepfake Detection: Generalisation Across Generation Methods Evaluating whether deepfake detection models trained on GAN-generated faces generalise to diffusion-model-generated faces, and investigating detection features robust across generation paradigms. Research Question: Do deepfake detection models trained exclusively on GAN-generated face datasets exhibit significant performance degradation when evaluated on diffusion-model-generated faces, and which low-level artifact features generalise across generation methods?
  Graduate
• 29
  Formal Verification of Security Protocols Using Model Checking Applying formal model checking tools (ProVerif, Tamarin) to verify the security properties of modern authentication protocols including OAuth 2.0 extensions and FIDO2/WebAuthn under Dolev-Yao attacker assumptions. Research Question: Does Tamarin model checking of the FIDO2 authentication protocol reveal any security property violations (authentication, anonymity, unlinkability) under standard Dolev-Yao adversary assumptions that are not exposed by informal analysis?
  PhD
• 30
  Privacy Leakage in Large Language Model Memorisation Investigating the extent to which LLMs memorise and regurgitate personally identifiable information from training data, and evaluating membership inference and data extraction attack success rates on open-weight models. Research Question: What fraction of verbatim training data sequences can be extracted from open-weight LLMs using targeted prompt strategies, and does model scale positively correlate with verbatim memorisation prevalence?
  Graduate

Data science research spans the boundary between computer science and statistics, addressing problems in data collection, storage, processing, analysis, and the extraction of actionable insight from large-scale, heterogeneous datasets. Research in this domain encompasses database systems (query optimisation, transaction management, NewSQL), data integration and quality, graph databases and knowledge graphs, stream processing, and the emerging intersection of data management with machine learning — a research area often called “ML for systems” and “systems for ML.” Strong data science research combines technical depth (new algorithms or systems with empirical evaluation) with domain knowledge of the application context.

• 31
  Automated Data Cleaning with Active Learning for Noisy Tabular Datasets Developing an automated pipeline that combines rule-based error detection with active learning to prioritise human annotation effort for data quality improvement in large tabular datasets with diverse error types. Research Question: Can active learning-guided data cleaning achieve equivalent data quality improvement with fewer human annotation labels than passive random sampling across real-world financial and healthcare tabular datasets?
  Graduate
• 32
  Knowledge Graph Completion Using Relational Graph Neural Networks Applying and extending relational GNN architectures for predicting missing links in large-scale knowledge graphs, benchmarking on Freebase and Wikidata subsets across relation type categories. Research Question: Does compositional relational path reasoning in R-GCN architectures outperform translation-based embedding models (TransE, RotatE) for knowledge graph completion on rare and long-tail relation types?
  Graduate
• 33
  Real-Time Anomaly Detection in Time-Series IoT Sensor Streams Designing lightweight anomaly detection algorithms for continuous IoT sensor streams that maintain accuracy with minimal computational overhead, evaluated on SWAT and MSL/SMAP benchmark datasets. Research Question: Do transformer-based time-series anomaly detection models outperform LSTM autoencoders on multivariate IoT sensor streams when both are constrained to real-time inference latency requirements?
  Undergrad
• 34
  Bias in Large-Scale Web-Crawled Training Datasets Characterising demographic, geographic, and linguistic biases in Common Crawl-derived datasets used to train foundation models, and measuring how dataset composition affects downstream model behaviour. Research Question: How does the geographic and linguistic composition of Common Crawl-derived pre-training corpora correlate with LLM performance disparities on culturally diverse benchmark tasks across world regions?
  PhD / MS
• 35
  Learned Query Optimisation in Relational Database Systems Replacing or augmenting traditional cost-based query optimisers with learned models that predict query execution costs from query structure and database statistics, evaluated on TPC-H and JOB benchmarks. Research Question: Do learned query optimisers trained on workload-specific execution traces achieve lower end-to-end query latency than PostgreSQL’s native cost-based optimiser on complex multi-join analytical queries?
  PhD
• 36
  Data Provenance Tracking for Machine Learning Pipelines Designing provenance tracking systems for ML pipelines that record data lineage from raw collection through preprocessing and model training, enabling debugging, auditability, and regulatory compliance. Research Question: Can automatic dataflow-based provenance capture for ML training pipelines enable retroactive root-cause analysis of model degradation events with acceptable storage and runtime overhead?
  Graduate
• 37
  Geospatial Big Data Analytics for Urban Mobility Pattern Discovery Applying scalable clustering and trajectory analysis algorithms to large-scale GPS and mobile phone location datasets to discover urban mobility patterns, with applications to transportation planning. Research Question: Do mobility pattern clusters derived from anonymised mobile network signalling data align sufficiently with ground-truth travel survey data to reliably inform public transport route planning decisions?
  Undergrad
• 38
  Graph-Based Recommendation Systems: Social Network Integration Incorporating social graph structure into collaborative filtering recommendation models to improve recommendation quality for cold-start users with limited interaction history. Research Question: Does integrating social network graph structure into graph attention network-based collaborative filtering significantly improve recommendation accuracy for cold-start users on the Amazon and Yelp benchmark datasets?
  Undergrad
• 39
  Stream Processing Latency and Fault Tolerance in Apache Flink vs. Kafka Streams Empirical comparison of Apache Flink and Kafka Streams for real-time stateful stream processing under high-throughput, failure injection, and late-arrival event scenarios. Research Question: Under equivalent cluster configurations and workload profiles, which stream processing framework achieves lower end-to-end latency for stateful aggregations while maintaining exactly-once processing guarantees after simulated node failures?
  Graduate
• 40
  Causal Inference Methods for Healthcare Observational Data Applying and comparing propensity score matching, instrumental variables, and doubly robust estimation for estimating treatment effects from electronic health record data in the absence of randomised controlled trials. Research Question: Do causal inference estimates of treatment effect derived from EHR observational data using doubly robust estimation converge with RCT results for comparable patient populations, and which confounding structures most threaten their validity?
  PhD
• 41
  Explainable Anomaly Detection in Financial Transaction Networks Building graph-based fraud detection systems for financial transaction networks that produce human-interpretable explanations alongside predictions, evaluated on the IEEE-CIS and Elliptic Bitcoin datasets. Research Question: Do GNN-based fraud detection systems with built-in attention-based explanations achieve both higher AUC-ROC and higher explanation fidelity on financial transaction graphs compared to post-hoc explainability applied to black-box classifiers?
  Graduate
• 42
  Multimodal Sentiment Analysis of Social Media Data Combining text, image, and video modalities for fine-grained sentiment classification on social media posts, evaluating fusion architectures and cross-modal attention mechanisms. Research Question: Does early fusion of text and visual modalities in transformer-based multimodal sentiment classifiers outperform late fusion on datasets where text and image sentiment signals conflict?
  Undergrad
• 43
  Data Mesh Architectures: Decentralised Data Ownership at Scale Evaluating the operational and quality trade-offs of data mesh architectures versus centralised data lake designs in large organisations, including domain ownership, schema governance, and data quality enforcement. Research Question: Does domain-oriented decentralised data ownership in data mesh architectures reduce time-to-insight for data consumers while maintaining data quality standards comparable to centralised data warehouse governance?
  Graduate

• 44
  Serverless Computing Cold Start Latency: Optimisation Strategies Characterising cold start latency in major FaaS platforms (AWS Lambda, Google Cloud Functions) and evaluating pre-warming, container snapshot, and WebAssembly-based strategies for latency reduction. Research Question: Do WebAssembly-based serverless function execution environments achieve lower cold start latency than container-based environments for latency-sensitive workloads without significant throughput penalties?
  Graduate
• 45
  QUIC Protocol Performance vs. TCP for Web Application Delivery Benchmarking QUIC and HTTP/3 against TCP and HTTP/2 for web content delivery under varying network conditions including packet loss, high latency, and mobile handoffs. Research Question: Does HTTP/3 over QUIC provide statistically significant page load time improvements over HTTP/2 for web applications with large resource counts under real-world packet loss rates?
  Undergrad
• 46
  Consensus Algorithm Performance in Geo-Distributed Blockchain Networks Comparing the latency, throughput, and fault tolerance of Raft, PBFT, and Tendermint consensus mechanisms under geo-distributed network configurations with simulated wide-area network latency. Research Question: How does increasing geographic distribution of consensus nodes affect the latency-throughput trade-off of PBFT vs. Tendermint Byzantine fault-tolerant consensus, and at what network size does Tendermint’s performance advantage emerge?
  Graduate
• 47
  Edge Computing Resource Allocation for Real-Time Video Analytics Designing and evaluating dynamic resource allocation algorithms for edge computing nodes running real-time video analytics workloads, optimising for latency under fluctuating workload and network conditions. Research Question: Can DRL-based resource allocation at edge nodes reduce inference latency for real-time video analytics below cloud offloading thresholds while maintaining accuracy of object detection at or above 90% mAP?
  PhD / MS
• 48
  5G Network Slicing for Industrial IoT Quality of Service Designing and simulating 5G network slice configurations for industrial IoT use cases with ultra-low latency requirements, evaluating slice isolation, admission control, and SLA compliance. Research Question: Does dynamic network slicing in 5G networks reliably meet the sub-millisecond latency and ultra-high reliability requirements of industrial automation applications under co-existing eMBB and mMTC slice traffic?
  PhD
• 49
  Kubernetes Autoscaling Strategies for Microservice Workloads Comparing horizontal pod autoscaling (HPA), vertical pod autoscaling (VPA), and predictive scaling approaches for microservice applications under realistic traffic patterns with varying burstiness. Research Question: Does predictive autoscaling using LSTM-based traffic forecasting reduce SLA violations and resource over-provisioning costs in Kubernetes clusters compared to reactive HPA for bursty microservice workloads?
  Graduate
• 50
  Distributed Caching Consistency: Redis vs. Memcached Under Write-Heavy Workloads Empirical evaluation of distributed caching systems under write-intensive workloads with cache invalidation requirements, measuring consistency guarantees, throughput, and latency under network partition simulation. Research Question: Under write-heavy workloads requiring strong cache consistency, does Redis Cluster’s consistency model result in measurably lower hit rates or higher latency compared to application-level cache invalidation strategies?
  Undergrad
• 51
  Software-Defined Networking for Dynamic DDoS Mitigation Designing and evaluating SDN controller applications that detect and mitigate volumetric DDoS attacks by dynamically reprogramming flow tables and rerouting traffic at line speed. Research Question: Can SDN-based DDoS mitigation using entropy-based traffic analysis reroute attack traffic and restore victim service availability within 5 seconds of attack onset without disrupting legitimate traffic flows?
  Graduate
• 52
  Byzantine Fault Tolerance in Distributed Machine Learning Analysing the vulnerability of gradient aggregation in distributed ML training to Byzantine worker attacks (gradient poisoning) and evaluating robust aggregation rules including Krum, Bulyan, and coordinate-wise median. Research Question: Do Byzantine-robust aggregation rules for distributed SGD maintain convergence and model accuracy under adversarial gradient manipulation when the fraction of Byzantine workers exceeds theoretical tolerance bounds?
  PhD
• 53
  Network Traffic Classification Using Encrypted Traffic Analysis Classifying application type and user activity from TLS-encrypted network flows using machine learning on metadata features (packet length distributions, inter-arrival times, flow duration) without decryption. Research Question: Can flow-level metadata features from TLS 1.3 encrypted traffic reliably classify application type with accuracy comparable to deep packet inspection, using only packet timing and size distributions?
  Undergrad
• 54
  Green Cloud Computing: Carbon-Aware Task Scheduling Designing cloud task scheduling algorithms that shift flexible computational workloads temporally and geographically to align with low-carbon electricity availability, evaluating against carbon and cost objective functions. Research Question: Can carbon-aware temporal task scheduling using real-time grid carbon intensity data reduce the carbon footprint of cloud batch workloads by more than 20% without exceeding user-defined deadline SLAs?
  Graduate
• 55
  P4-Programmable Data Planes for In-Network Machine Learning Inference Implementing simplified machine learning classifiers directly on P4-programmable network switches for line-rate traffic classification, evaluating model fidelity versus latency trade-offs. Research Question: Can decision tree classifiers implemented on P4-programmable network switches achieve sufficient classification accuracy for network traffic anomaly detection at line rate (100 Gbps) with acceptable memory constraints?
  PhD
• 56
  Satellite Network Integration for Low-Latency Global Internet Access Analysing the performance characteristics of low-Earth-orbit satellite constellations (Starlink, OneWeb) for latency-sensitive applications and evaluating routing optimisation strategies for ground-LEO-ground paths. Research Question: Do LEO satellite constellations achieve comparable round-trip latency to terrestrial fibre for trans-oceanic paths, and what routing algorithms minimise handoff-induced latency spikes in mobility scenarios?
  Graduate

• 57
  AI Code Generation Quality: Correctness and Security of LLM-Generated Code Systematically evaluating the correctness, security, and maintainability of code generated by LLMs (GitHub Copilot, GPT-4, Claude) across programming tasks, measuring vulnerability prevalence and functional correctness rates. Research Question: What fraction of LLM-generated code completions introduce security vulnerabilities (OWASP Top 10 categories) compared to human-written code on equivalent tasks, and does larger model scale correlate with reduced vulnerability rate?
  Graduate
• 58
  Automated Test Case Generation Using Large Language Models Evaluating LLMs for automated unit test generation — measuring code coverage, mutation score, and false positive rates compared to EvoSuite and Randoop on open-source Java projects. Research Question: Do LLM-generated unit tests achieve mutation scores comparable to EvoSuite-generated tests on real open-source Java projects, and which code complexity characteristics predict LLM test generation quality?
  Graduate
• 59
  Technical Debt Quantification Using Static Analysis Metrics Developing and validating metrics for quantifying technical debt in large codebases using static analysis, and measuring the correlation between technical debt indicators and maintenance cost in industrial software systems. Research Question: Do static analysis-derived technical debt indicators (cyclomatic complexity, code churn, code smells) reliably predict bug density and maintenance effort in open-source Java and Python projects?
  Undergrad
• 60
  Mutation Testing Effectiveness: Are More Mutants Better? Investigating whether larger mutant sets produce more reliable test suite quality assessments than subsampled mutation sets, and evaluating higher-order mutation operators for revealing complex bugs. Research Question: Does reducing mutation set size through subsampling strategies preserve the ranking of test suite quality assessments produced by full mutation analysis, and what minimum sample size maintains ranking fidelity?
  PhD
• 61
  Microservices vs. Monolithic Architecture: Maintainability at Scale Empirically studying whether microservices architectures exhibit lower change failure rates, shorter recovery times, and lower deployment frequency inhibitors than equivalent monolithic systems in mature software organisations. Research Question: Do software teams operating microservice architectures achieve statistically better DORA metrics (deployment frequency, lead time, change failure rate, MTTR) than teams operating monolithic architectures of comparable business complexity?
  Undergrad
• 62
  Continuous Integration Pipeline Flakiness: Root Causes and Remediation Analysing the prevalence, root causes, and developer cost of flaky tests in large-scale CI/CD pipelines, and evaluating automated flakiness detection and quarantine strategies. Research Question: What categories of flakiness root causes (timing dependencies, order dependencies, resource leaks, environmental dependencies) are most prevalent in CI pipelines of large-scale open-source projects, and which categories are most remediable by automated approaches?
  Graduate
• 63
  Formal Specification and Verification of Safety-Critical Software Applying model checking and theorem proving tools (TLA+, Coq, Isabelle/HOL) to specify and verify safety properties of aviation or automotive control software components. Research Question: Can TLA+ model checking of distributed consensus components in aviation communication systems discover safety property violations that are not detected by comprehensive simulation-based testing?
  PhD
• 64
  Code Review Automation: Prioritising Review Comments by Severity Building and evaluating ML models that predict the importance and actionability of code review comments, enabling automated prioritisation of review attention for high-severity issues. Research Question: Can transformer-based models trained on historical code review comment datasets predict reviewer-assessed comment severity with sufficient accuracy to meaningfully reduce review time spent on low-severity feedback?
  Undergrad
• 65
  Agile Estimation Accuracy: Story Points vs. Probabilistic Forecasting Comparing traditional story point estimation with probabilistic flow-based forecasting methods (Monte Carlo simulation over historical throughput) for software development effort estimation accuracy. Research Question: Does probabilistic Monte Carlo forecasting using team throughput data produce more accurate sprint delivery predictions than story point estimation across diverse Agile team contexts and project types?
  Undergrad
• 66
  Developer Experience and Cognitive Load in Modern Software Toolchains Measuring the cognitive load and context-switching costs imposed by complex modern software toolchains on developer productivity, using physiological measures and code complexity metrics alongside self-reported experience surveys. Research Question: Does toolchain complexity — measured by the number of distinct tools, configuration files, and context switches required per development task — correlate with increased error rates and reduced task completion speed in controlled developer studies?
  Graduate
• 67
  Green Software Engineering: Energy Consumption of Programming Language Runtimes Benchmarking the energy consumption of equivalent algorithmic implementations across programming languages and runtimes, identifying the computational patterns that drive efficiency disparities. Research Question: Do compiled systems languages (Rust, C++) consistently achieve lower energy consumption than JVM-based (Java, Kotlin) and interpreted (Python, JavaScript) runtimes across a diverse set of computational problem categories on equivalent hardware?
  Undergrad
• 68
  Differential Fuzzing for API Compatibility Testing Across Library Versions Applying differential fuzzing to detect breaking behaviour changes between consecutive versions of widely-used software libraries, automatically identifying regression-inducing API changes. Research Question: Can automated differential fuzzing between adjacent library versions detect API behavioural regressions that are not captured by existing regression test suites in popular open-source Python and Java libraries?
  Graduate

• 69
  Conversational AI Trust Calibration: When Users Over-Trust Chatbots Studying the conditions under which users over-trust conversational AI systems for high-stakes decisions (medical, legal, financial), and evaluating interface interventions designed to promote appropriate trust calibration. Research Question: Do uncertainty disclosure mechanisms in conversational AI interfaces — expressing model confidence as natural language hedging — significantly reduce inappropriate high-stakes decision deference compared to interfaces without uncertainty disclosure?
  Graduate
• 70
  Accessibility of AI-Generated Content for Screen Reader Users Evaluating the accessibility of AI-generated text, images, and data visualisations for blind and low-vision users using screen readers, and proposing accessible-by-design generation strategies. Research Question: Do existing alt-text generation approaches for AI-generated images meet WCAG 2.1 accessibility standards as assessed by blind users’ task completion rates and subjective comprehension, and which image types present the greatest accessibility gaps?
  Undergrad
• 71
  Dark Patterns in Mobile Applications: Detection and Regulatory Implications Systematically cataloguing deceptive UI design patterns in mobile applications from major app stores, measuring their prevalence across categories, and assessing alignment with emerging EU Digital Services Act compliance requirements. Research Question: What categories of dark patterns are most prevalent in top-grossing mobile applications, and do applications from different regulatory jurisdictions differ significantly in their use of deceptive interface design?
  Undergrad
• 72
  Cognitive Load in Multi-Tab Browser Workflows: Eye-Tracking Evidence Using eye-tracking and think-aloud protocols to measure the cognitive load imposed by multi-tab browsing on information retrieval tasks, and evaluating tab management interface designs that reduce working memory burden. Research Question: Does multi-tab browser usage during complex information retrieval tasks produce measurable increases in cognitive load (fixation duration, saccade frequency) compared to single-context navigation, and do AI-assisted tab grouping features reduce this load?
  Graduate
• 73
  Voice User Interface Design for Older Adults with Mild Cognitive Impairment Designing and evaluating voice user interface interaction patterns and prompting strategies for smart home control by older adults with mild cognitive impairment, measuring task completion, error recovery, and user experience. Research Question: Do confirmatory feedback and re-prompting strategies in voice interfaces for smart home control significantly improve task completion rates for older adults with mild cognitive impairment compared to standard VUI interaction models?
  Graduate
• 74
  User Mental Models of Machine Learning: Misconceptions and Calibration Investigating the mental models non-expert users form about how ML-based systems (recommendation, prediction, content moderation) make decisions, and evaluating explanation interfaces that correct common misconceptions. Research Question: What systematic misconceptions about machine learning decision-making do non-expert users hold when interacting with ML-based content recommendation, and do feature-importance explanations effectively correct these misconceptions?
  Undergrad
• 75
  Mixed Reality Interfaces for Remote Collaborative Engineering Design Developing and evaluating mixed reality collaboration platforms for distributed engineering design teams, measuring task performance, situation awareness, and communication quality compared to video conferencing baselines. Research Question: Do mixed reality collaboration platforms providing shared 3D design artefact manipulation improve engineering design task performance and team coordination compared to video conferencing for geographically distributed teams?
  PhD
• 76
  Inclusive Design of Generative AI Creative Tools Evaluating whether generative AI image and text creation tools are equally usable and produce equally satisfactory outputs for users with different levels of creative experience, disabilities, and linguistic backgrounds. Research Question: Do users with visual creative experience exhibit significantly different satisfaction and output quality when using text-to-image generative AI tools compared to users without visual training, and what interface adaptations reduce this experience gap?
  Undergrad
• 77
  Notification Design and Interruption Management in Productivity Applications Studying the relationship between notification timing, urgency signalling, and interruption recovery cost in knowledge worker productivity applications, and evaluating intelligent notification batching designs. Research Question: Does AI-based notification timing optimisation that defers non-urgent notifications to task breakpoints reduce disruption cost — measured as time to return to pre-interruption performance — without increasing perceived urgency concerns?
  Graduate
• 78
  Algorithmic Transparency Requirements: User Comprehension and Behaviour Change Testing whether regulatory-mandated algorithm transparency disclosures (GDPR Article 22 explanations, EU AI Act transparency requirements) actually enable users to understand and contest automated decisions in practice. Research Question: Do GDPR-mandated algorithmic decision explanations provide sufficient information for non-expert users to identify and contest incorrect automated credit or hiring decisions, and what explanation formats maximise comprehension?
  Graduate
• 79
  Attention Management in Autonomous Vehicle Interfaces Designing and evaluating takeover request interfaces for Level 3 autonomous vehicles that optimise driver re-engagement speed and accuracy when manual control is required from an automated driving state. Research Question: Does multimodal takeover request design (visual + auditory + haptic) in Level 3 autonomous vehicles achieve faster and safer driver takeover responses than unimodal requests while avoiding startle responses that degrade control quality?
  PhD

• 80
  Approximation Algorithms for the Capacitated Facility Location Problem Designing and analysing constant-factor approximation algorithms for capacitated variants of the facility location problem, with applications to supply chain and cloud data centre placement optimisation. Research Question: Can LP-rounding techniques for the soft-capacitated facility location problem achieve an approximation ratio below 3 when client demands are non-uniform and facility capacities are heterogeneous?
  PhD
• 81
  Online Learning Algorithms for Dynamic Pricing Under Demand Uncertainty Designing online algorithms with provable regret bounds for revenue maximisation pricing problems where demand distribution is unknown and must be learned through pricing decisions. Research Question: Can online learning algorithms for dynamic pricing achieve sublinear regret bounds under non-stationary demand when equipped with change-point detection, and how does detection latency affect the regret-exploration trade-off?
  PhD
• 82
  Subgraph Isomorphism Detection: Practical Algorithms for Large Graphs Engineering and benchmarking practical subgraph isomorphism detection algorithms on large real-world graphs (social networks, protein interaction networks), comparing VF2++, RI, and backtracking variants. Research Question: What graph structural properties (density, degree distribution, diameter) most strongly determine the relative performance of VF2++ versus RI for subgraph isomorphism detection on real-world biological and social network graphs?
  Graduate
• 83
  Cache-Efficient Algorithms for Processing Large Sparse Matrices Designing and benchmarking sparse matrix storage formats and traversal algorithms optimised for modern cache hierarchies, targeting graph analytics and scientific computing workloads. Research Question: Do cache-blocked sparse matrix-vector multiplication algorithms achieve consistent speedups over CSR-format implementations across sparse matrices with varying structural characteristics (density, row distribution, sparsity pattern)?
  Graduate
• 84
  Differential Privacy Mechanisms for Graph Data Analysis Designing differentially private algorithms for graph statistics (degree distribution, clustering coefficient, triangle count) that provide strong privacy guarantees while maintaining utility for downstream analytics. Research Question: Can edge-differential privacy mechanisms for triangle counting in social graphs achieve utility within 10% of non-private baseline estimates at privacy budgets (ε < 1) acceptable for practical deployment?
  PhD
• 85
  Randomised Algorithms for Streaming Cardinality Estimation Analysing and extending HyperLogLog and related sketch-based cardinality estimation algorithms for high-cardinality data streams, evaluating accuracy, memory, and update throughput trade-offs. Research Question: Do HyperLogLog++ modifications with adaptive correction functions achieve lower relative error than standard HyperLogLog at equivalent memory budgets across streams with different element distributions?
  Graduate
• 86
  Fixed-Parameter Tractability for Network Interdiction Problems Investigating parameterised complexity of graph network interdiction problems — identifying parameters (treewidth, solution size, planarity) that enable FPT algorithms for NP-hard interdiction variants. Research Question: Is the minimum-cost network interdiction problem fixed-parameter tractable when parameterised by the treewidth of the underlying graph, and does an FPT algorithm yield practical speedups over ILP formulations on infrastructure network instances?
  PhD
• 87
  Parallel Sorting Algorithm Performance on Modern Many-Core Processors Benchmarking parallel sorting algorithm variants (parallel mergesort, radix sort, bitonic sort) on GPU and multi-core CPU architectures for different input distributions and sizes. Research Question: Which parallel sorting algorithm achieves the most consistent throughput across diverse input distributions (uniform, Zipfian, nearly sorted) on modern GPU architectures with high thread counts?
  Undergrad
• 88
  Algorithmic Game Theory: Mechanism Design for Fair Resource Allocation Designing strategyproof and envy-free resource allocation mechanisms for divisible goods under heterogeneous agent preferences, with applications to bandwidth allocation and cloud computing markets. Research Question: Is simultaneous strategyproofness and envy-freeness achievable for divisible resource allocation with heterogeneous agent preferences beyond two-agent settings, and if not, which property is more efficiently approximated?
  PhD
• 89
  String Matching Algorithms for Compressed Text: Searching Without Decompression Designing and benchmarking algorithms for substring search directly on compressed text representations (BWT, LZ77, grammar-compressed), evaluating the time-space trade-off versus decompress-then-search baselines. Research Question: Do suffix-array-based compressed text index structures provide practical search performance advantages over decompress-then-search approaches for genome sequence databases with high repetitiveness at genome-scale data sizes?
  Graduate
• 90
  Lower Bounds for Data Structure Problems via Cell-Probe Complexity Investigating cell-probe complexity lower bounds for fundamental data structure problems including dynamic connectivity, predecessor search, and partial sums, extending the polynomial method framework. Research Question: Can cell-probe lower bounds for dynamic connectivity be strengthened using the polynomial method beyond existing Ω(log n / log log n) bounds for the amortised operation cost?
  PhD
• 91
  Compiler Optimisation for Tensor Operations in Deep Learning Frameworks Evaluating compiler-level loop tiling, operator fusion, and memory layout optimisation strategies for tensor computations in PyTorch and JAX, targeting both training and inference performance. Research Question: Do polyhedral model-based compiler optimisations for tensor operation scheduling in deep learning frameworks achieve consistent training throughput improvements over auto-tuned handwritten CUDA kernels across model architecture families?
  Graduate

• 92
  Quantum Error Correction: Surface Code Implementation on NISQ Devices Implementing and characterising surface code error correction on near-term noisy intermediate-scale quantum (NISQ) hardware, measuring logical error rates versus physical qubit overhead. Research Question: What is the minimum code distance at which surface code error correction on current superconducting qubit hardware achieves a logical error rate below the physical error rate, and how does crosstalk noise affect this threshold?
  PhD
• 93
  Variational Quantum Algorithms for Combinatorial Optimisation Implementing and benchmarking QAOA and VQE for combinatorial optimisation problems (Max-Cut, travelling salesman), evaluating whether quantum advantage is achievable at current hardware scales. Research Question: Does QAOA at practical circuit depths (p ≤ 10) on current quantum hardware achieve better approximation ratios for Max-Cut than classical Goemans-Williamson on problem instances where quantum advantage is theoretically expected?
  PhD / MS
• 94
  Blockchain Scalability: Layer-2 Solutions for Ethereum — Optimistic vs. ZK Rollups Comparing optimistic rollup and zero-knowledge rollup approaches for Ethereum scalability, measuring transaction throughput, finality time, proof generation cost, and security assumptions. Research Question: Do ZK-rollup solutions provide meaningful finality time advantages over optimistic rollups for DeFi applications on Ethereum, and at what transaction volume does the proof generation cost of ZK-rollups become economically competitive?
  Graduate
• 95
  Neuromorphic Computing for Always-On Edge Inference Evaluating neuromorphic chips (Intel Loihi 2, IBM NorthPole) for spiking neural network inference on sensor data classification tasks, measuring energy efficiency versus accuracy compared to conventional edge AI accelerators. Research Question: Do spiking neural network implementations on neuromorphic hardware achieve energy efficiency advantages over equivalent accuracy conventional neural networks on inference accelerators for sparse, event-driven sensor classification tasks?
  PhD
• 96
  Digital Twin Security: Attack Surfaces in Cyber-Physical System Twins Analysing the attack surfaces introduced by industrial digital twin deployments, including synchronisation protocol vulnerabilities, inference attacks on twin models, and spoofing of physical sensor data. Research Question: What novel attack vectors do industrial digital twin deployments introduce beyond the security threat model of the underlying cyber-physical system, and which are exploitable through the twin’s network interface without physical access?
  Graduate
• 97
  Decentralised Identity and Self-Sovereign Identity Systems Designing and evaluating decentralised identity management systems based on W3C DID and verifiable credential standards, measuring privacy properties, revocation performance, and user adoption friction. Research Question: Do W3C DID-based self-sovereign identity systems provide sufficient credential revocation performance and privacy properties for healthcare patient identity management at national health system scales?
  Graduate
• 98
  In-Memory Computing for Transformer Attention Computation Investigating the use of resistive RAM (ReRAM) and phase-change memory (PCM) crossbar arrays for performing matrix-vector multiplication in transformer attention layers, evaluating energy and latency advantages. Research Question: Can in-memory computing using resistive crossbar arrays perform transformer attention head computation with sufficient numerical precision to maintain model accuracy while achieving meaningful energy reduction over SRAM-based inference?
  PhD
• 99
  Extended Reality for STEM Education: Learning Outcomes vs. Traditional Methods Evaluating the learning outcomes of XR-based interactive simulations for university-level STEM education compared to traditional laboratory and lecture methods, across physics, chemistry, and biology domains. Research Question: Do XR-based laboratory simulations for university chemistry education produce equivalent conceptual understanding and procedural skill outcomes as physical lab practicals, and for which experiment categories are XR simulations most effective?
  Undergrad
• 100
  Federated Analytics for Distributed IoT Fleet Management Designing privacy-preserving federated analytics systems for industrial IoT fleets that compute aggregate statistics over device populations without centralising raw sensor data or individual device telemetry. Research Question: Can federated analytics with local differential privacy guarantees provide fleet-level predictive maintenance accuracy comparable to centralised telemetry analysis, at acceptable privacy budgets for industrial IoT deployments?
  Graduate
• 101
  Carbon Footprint of Generative AI: Lifecycle Energy Analysis Measuring and modelling the full lifecycle energy consumption and carbon footprint of large generative AI systems — from pre-training through fine-tuning, inference serving, and model retirement — across different infrastructure configurations. Research Question: What fraction of a generative AI system’s total lifecycle carbon footprint is attributable to inference serving versus pre-training, and how does this ratio change with model size, query volume, and geographical data centre carbon intensity?
  Graduate
• 102
  Homomorphic Encryption for Secure Cloud Computation Benchmarking CKKS and BFV fully homomorphic encryption schemes for practical machine learning inference on encrypted data, measuring computational overhead and noise accumulation across model types. Research Question: Is CKKS-scheme homomorphic encryption computationally feasible for real-time encrypted inference of convolutional neural network models on cloud infrastructure, and what model depth and accuracy thresholds are achievable within practical latency constraints?
  PhD / MS
• 103
  Quantum Machine Learning: Advantage Claims and Dequantisation Critically evaluating proposed quantum speedups for machine learning algorithms, investigating dequantisation results that provide classical algorithms with comparable complexity, and characterising the conditions for genuine quantum advantage. Research Question: For which machine learning problem categories do quantum algorithms provide provable speedups that are not achievable by classical sample-based algorithms, given access to equivalent data via quantum-RAM versus classical sampling oracles?
  PhD

How to Choose Your Computer Science Research Topic

The 103 topics above represent a curated selection across eight domains — but choosing among them, or using them as inspiration for your own original topic, requires a structured process. The most common mistake student researchers make is choosing a topic they find intellectually interesting without verifying that it is both novel (the research gap genuinely exists) and feasible (the question can be answered with available resources within the available time). Both checks are essential, and both must be completed before serious investment in a topic begins.

Computer Science Research Methodology: From Question to Publishable Result

Computer science research is methodologically diverse — spanning formal proofs, simulation, empirical system evaluation, user studies, and surveys. The choice of methodology must be driven by the type of research question being asked. A design question (can I build a system that does X?) requires system implementation and empirical evaluation against baselines. A measurement question (how does existing system Y behave?) requires controlled experimental design and statistical analysis. A theoretical question (can I prove that algorithm Z has property W?) requires formal mathematical argument. Understanding this alignment between question type and methodology is one of the most important intellectual skills in CS research.

Common Mistakes in Computer Science Research Papers

Strong vs. Weak Research Claims

Pre-Submission Research Checklist

• Research question is precise, specific, and answerable — it names method, comparison, metric, and experimental setting
• All prior work directly relevant to your contribution is cited, discussed, and fairly compared
• Baselines are state-of-the-art, carefully implemented, and evaluated under equivalent conditions to your method
• Ablation studies isolate the contribution of each proposed component
• All quantitative results include measures of uncertainty (standard deviation, confidence intervals, p-values)
• Experiments are reproducible: code released, hyperparameters specified, hardware documented, seeds fixed
• Limitations are discussed honestly — including conditions where your method does not outperform baselines
• Ethical implications of the research (dual-use risks, data privacy, potential harms) are explicitly addressed
• The paper is targeted at an appropriate venue for its contribution type and expected impact level
• Figures and tables communicate results clearly, with axis labels, units, and error bars where applicable

FAQs: Computer Science Research Topics Answered

Conclusion: The Right Research Topic Is the Beginning of a Contribution

The 103 computer science research topics in this guide represent a cross-section of the field at a single point in time — July 2025 — in a discipline that moves faster than almost any other in academia. Six months from now, some of the “emerging technology” topics in this guide will be mainstream; new sub-fields will have emerged from the intersection of quantum computing and machine learning, or from the practical challenges of deploying LLMs at planetary scale, or from the security implications of ubiquitous AI decision-making in critical infrastructure. Staying current requires not just reading this guide but maintaining a habit of regular literature monitoring — following key researchers on Semantic Scholar, reading arXiv preprints in your area, and attending or watching recordings of top conference proceedings.

What will not change is the fundamental structure of what makes computer science research excellent: a precise and novel research question, a methodology rigorously designed to answer it, an honest and statistically valid evaluation, and writing that communicates the contribution with enough clarity that other researchers can reproduce, build on, and challenge your work. The specific research questions change with the technology; the intellectual virtues of good research are permanent.

Whether you are selecting a topic for your first undergraduate research project or narrowing a dissertation focus after years of graduate study, the process is the same: find something genuinely important that is not yet understood, design a rigorous way to understand it, and communicate what you discover as clearly and honestly as possible. The topics in this guide are starting points for that process — not destinations, but the first step on a path that, when followed with intellectual rigour and genuine curiosity, can produce work that genuinely advances human knowledge.

For expert support with your computer science research at any stage — from topic selection and literature review through implementation, evaluation, and final paper writing — visit Smart Academic Writing and explore our computer science assignment help, research paper writing, dissertation writing, data analysis support, and editing and proofreading services.