Topics for Capstone Project in Information Technology

We know capstone projects open doors when they ship real value and clear results. Today we at TopicSuggestions share practical, high-impact IT topics shaped by what we track in industry and what we see in successful student builds each term. We recognize that the hardest step is choosing a scope that fits a semester, matches your skills, and has data or users you can actually reach. We aim to give you a focused list of buildable ideas you can defend, implement, and showcase.

Information Technology Capstone Project Topic Ideas

We will group topics by area—AI and data, cybersecurity, web and mobile, cloud and DevOps, IoT and embedded, plus domain tracks like health tech and fintech—and for each topic we will note the problem statement, suggested stack, data or test plan, and a stretch goal so you can pick fast and plan with confidence.

1. Olfactory Semantics for NLP: Translating Electronic-Nose Streams into Descriptive Language

We ask how we can map high-dimensional volatile compound sensor streams to fine-grained, culture-specific scent descriptors?
We investigate whether cross-lingual scent lexicons can be induced from paired olfactory–text corpora without parallel data?
We examine how we can robustly handle sensor drift over time via continual domain adaptation in scent-to-text translation?
We test whether interactive, human-in-the-loop scent captioning improves downstream tasks like recipe retrieval and wine profiling?

2. Dream-Report Augmented Pretraining: Leveraging Nocturnal Narratives for Metaphor and Surreal Reasoning

We ask whether pretraining on dream reports improves metaphor resolution and counterfactual narrative consistency?
We probe how we can automatically segment and label oneiric structures (scene shifts, impossible physics) to supervise models?
We test if dream-augmented models hallucinate differently from baselines under stress-tested factual QA?
We investigate how we can transfer gains from dream corpora to creative writing assistance without amplifying confabulation?

3. Chronemic Pragmatics in Speech NLP: Treating Pauses as Learnable Tokens for Social Meaning

We ask whether representing sub-second pauses as discrete tokens improves sarcasm detection and turn-taking prediction?
We evaluate how we can adapt pause semantics across cultures with divergent chronemic norms?
We test whether pause-aware models are more robust to ASR jitter and latency distortions in real-time dialogue?
We investigate how we can calibrate uncertainty from pausing patterns to flag conversational breakdowns early?

4. Learning Physical Commonsense from Household Robot Failure Narratives

We ask how we can extract causal schemas from natural-language descriptions of robot mishaps to improve text-based reasoning?
We evaluate whether narrativized failures transfer to counterfactual question answering about everyday physics?
We test how we can align learned schemas with liability-aware explanations that are legally and ethically defensible?
We investigate whether human-authored vs. robot-autogenerated failure stories differentially shape model generalization?

5. Bioacoustic-to-Text for Botany: Translating Plant Ultrasonic Emissions into Human-Legible Stress Reports

We ask how we can learn mappings from plant ultrasound to textual descriptors of stress, hydration, and pathogen load?
We evaluate whether multimodal alignment with spectral imagery improves interpretability and calibration of the text output?
We test how we can design farmer-in-the-loop interfaces that summarize plant states without anthropomorphizing?
We investigate transfer across species and environments to enable early-warning narratives for precision agriculture?

6. Jurisdiction-Aware Negotiation Generation: Dialogue That Obeys Contract Law Across Regions

We ask how we can encode statutory and case-law constraints to generate enforceable, jurisdiction-specific negotiation turns?
We evaluate whether neural–symbolic validators can block illegal clauses while preserving persuasive strategy?
We test how we can support multi-jurisdiction bargaining where parties anchor in different legal regimes?
We investigate robustness to adversarial prompts that attempt to smuggle in unenforceable terms?

7. Collective Memory Editing in Multi-Agent LLM Societies

We ask how we can design protocols for agents to propose, vet, and revise shared memory without collapsing into echo chambers?
We evaluate metrics for stability, diversity, and misinformation resilience under adversarial memory injections?
We test whether role-specialized editors (fact-checker, ethicist, historian) improve long-horizon task performance?
We investigate how we can attribute responsibility for downstream errors to specific memory edits?

8. Geo-Embodied NLP: Aligning Urban Microclimate Sensor Arrays with Neighborhood Narratives

We ask how we can map street-level sensor traces (heat, humidity, particulates) to localized, multilingual risk briefings?
We evaluate whether language-grounded microclimate models improve heatwave preparedness messaging for vulnerable groups?
We test domain adaptation across cities with differing infrastructure and dialectal variation?
We investigate how we can quantify causal impact of text alerts on subsequent human mobility and exposure?

9. Holographic Telepresence NLP for Sign Languages: Real-Time Repartee and Politeness Strategies

We ask how we can integrate skeletal, handshape, and facial cues into dialogue systems that generate fluent signed repartee?
We evaluate whether culture-specific politeness norms in sign languages can be modeled for turn-taking and backchanneling?
We test how we can achieve low-latency translation between signed dialogue and spoken text in holographic meetings?
We investigate robustness to occlusions and camera angle shifts while preserving linguistic nuance?

10. Ethogram-to-Narrative Translation: Turning Animal Behavior Sequences into Field Reports

We ask how we can map ethogram event streams into concise, bias-aware textual summaries for ecologists?
We evaluate whether structure-aware decoders reduce anthropomorphism while retaining interpretability?
We test cross-species transfer from well-studied behaviors to scarce-data habitats?
We investigate human–AI coauthoring workflows that preserve scientific rigor in automatically generated field notes?

11. Federated Continual Learning for Battery-Constrained Aquaculture IoT

We (Team TopicSuggestions) propose creating a federated continual learning system that adapts models on battery-powered aquaculture sensors without central dataset transfers.
We ask whether on-device replay buffers and selective parameter freezing can maintain accuracy under strict energy budgets, how to schedule model updates across heterogeneous devices to maximize lifetime, and how to validate model drift detection with sparse connectivity.
We will prototype lightweight learners on microcontrollers, simulate federated rounds with variable connectivity, measure energy via emulation or real batteries, and compare accuracy and lifetime against centralized and classical federated baselines.

12. Explainable Reinforcement Learning for Adaptive Urban Traffic Signal Control

We (Team TopicSuggestions) propose an explainable RL architecture that generates human-interpretable rationales for signal changes while optimizing throughput.
We ask how to align saliency-based explanations with traffic engineer mental models, whether causal attribution methods improve trust without reducing performance, and how to evaluate explanations under rare events (accidents, large crowds).
We will train RL agents in realistic traffic simulators, integrate explanation modules (counterfactuals, causal graphs), conduct user studies with traffic operators, and measure trade-offs between transparency, safety, and delay reduction.

13. Quantum-Resilient Permissioned Blockchain for Supply-Chain Provenance

We (Team TopicSuggestions) propose a permissioned blockchain design that replaces vulnerable cryptographic primitives with quantum-resistant schemes while preserving throughput and auditability for supply chains.
We ask what hybrid signature and hashing strategies achieve practical performance on commodity servers, how key management can be decentralized for consortium members, and how to prove end-to-end provenance under post-quantum assumptions.
We will benchmark candidate lattice-based primitives in a Fabric-like network, design consensus and key-rotation protocols, and run supply-chain use-case experiments to evaluate latency, storage, and forensic guarantees.

14. EEG-Informed Adaptive UI for Neurodiverse Users in Learning Platforms

We (Team TopicSuggestions) propose an adaptive user interface that uses lightweight EEG signals to infer attention and sensory overload in neurodiverse learners and dynamically adjust presentation.
We ask which EEG features correlate reliably with task engagement across individuals, how to map those features to safe UI adaptations (pace, color, complexity), and how to ensure privacy and low-friction calibration.
We will collect short calibration sessions with neurodiverse participants using portable EEG, train subject-adaptive classifiers, implement UI adaptation rules in a web learning platform, and evaluate learning outcomes and user comfort in controlled trials.

15. Self-Healing Microservices via Causal Tracing and Reinforcement Policies

We (Team TopicSuggestions) propose a self-healing orchestrator that uses causal trace graphs and RL-based repair policies to diagnose and remediate microservice failures automatically.
We ask how to construct efficient causal graphs from distributed tracing data at scale, which repair actions should be learned versus rule-based, and how to avoid repair oscillations or unsafe rollbacks.
We will instrument a containerized microservice benchmark with tracing, synthesize fault injections, learn repair policies in simulation, and validate in staging clusters measuring MTTR, false positives, and stability.

16. Cross-Lingual, Privacy-Preserving Synthetic Health Data for Rare Diseases

We (Team TopicSuggestions) propose generating multilingual synthetic clinical narratives for rare diseases that preserve utility for NLP tasks while enforcing differential privacy across jurisdictions.
We ask how to balance utility for entity extraction and cohort discovery with rigorous privacy budgets, how to model culturally specific medical language across languages, and how to evaluate rare-event fidelity.
We will adapt multilingual generative models with DP-SGD, curate multilingual rare-disease corpora via partnerships, evaluate downstream task performance and privacy leakage, and produce guidelines for cross-border data sharing.

17. Edge-Assisted Augmented Reality for Remote Industrial Maintenance with Semantic Compression

We (Team TopicSuggestions) propose an edge-cloud AR pipeline that compresses scene semantics for remote experts to maintain situational awareness under constrained uplink.
We ask what task-aware semantic representations (parts, tool affordances) compress best for remote diagnosis, how to allocate processing between headset, edge, and cloud adaptively, and how to synchronize sketch-based instructions with live semantics.
We will develop semantic segmentation and object-affordance models optimized for mobile GPUs, design adaptive offloading policies, run field trials with technicians using simulated bandwidth constraints, and measure repair time and comprehension.

18. Energy-Aware Scheduling for Heterogeneous GPU Clusters Using Graph Neural Policies

We (Team TopicSuggestions) propose using graph neural networks to schedule ML workloads on heterogeneous GPU clusters to minimize energy while meeting SLAs.
We ask how to represent job and device heterogeneity as dynamic graphs, whether learned policies can generalize to unseen workloads, and how to incorporate thermal and power capping constraints safely.
We will build simulators fed by cluster traces, train GNN-based schedulers with RL to optimize energy-SLA trade-offs, validate policies on a testbed with mixed GPUs, and compare to heuristic and optimizer baselines.

19. Citizen-Sourced Digital Twin for Urban Flood Resilience with Uncertainty-Aware Fusion

We (Team TopicSuggestions) propose a digital twin that fuses citizen sensor reports, smartphone imagery, and municipal telemetry to model urban flooding with quantified uncertainty for real-time decision support.
We ask how to calibrate low-fidelity citizen reports against official gauges, which fusion architectures propagate uncertainty transparently, and how to present probabilistic forecasts to diverse stakeholders for evacuation decisions.
We will create data pipelines to ingest crowdsourced reports, develop Bayesian fusion models for hydrology-informed digital twins, run controlled flooding simulations, and evaluate forecast utility in tabletop emergency exercises.

20. Composable Security Policy Synthesis Using Formal Methods and Machine Learning

We (Team TopicSuggestions) propose synthesizing composable security policies for federated applications by combining formal specification templates with learned intent extraction from developer repositories.
We ask how to map high-level security intents to formally verifiable policy fragments, how to ensure composability across modules without emergent conflicts, and how to provide developer-facing recommendations that are both secure and practical.
We will mine code and policy corpora for intent-patterns, design a synthesis engine that outputs formally checkable policies (SMT-based), validate composability via model checking, and run developer studies to assess usability and deployment friction.

21. Adaptive Federated Learning Incentive Mechanisms for Energy-Constrained IoT Devices

We propose an incentive-aware federated learning protocol that adapts communication frequency and model aggregation to heterogeneous energy budgets.
We ask: How can we design incentives that balance model utility with per-device energy constraints?; How do adaptive aggregation schedules affect convergence and fairness?; How can we prove incentive compatibility when devices may misreport energy?
We will work by designing a mathematical incentive model, simulating heterogeneous IoT fleets, implementing a prototype on ESP32/ARM devices, and measuring energy, accuracy, and participation fairness under different incentive rules.

22. Explainable Adversarial Robustness for TinyML Models on Embedded Vision

We propose methods that combine lightweight explainability with adversarial defense specifically for TinyML image classifiers on microcontrollers.
We ask: Which explanation techniques reveal the most attack-relevant features on constrained models?; How can explanations be used to detect and mitigate adversarial inputs in real time on-device?; What is the trade-off between explainability overhead and robustness gain?
We will work by training TinyML models, developing compact saliency/explanation modules, generating on-device adversarial examples, and evaluating detection rates, latency, and memory footprint on representative hardware.

23. Quantum-Inspired Lightweight Key Management for Fog Computing Nodes

We propose a quantum-inspired, classical-key-management scheme tailored to fog nodes that balances forward secrecy and low computational cost.
We ask: Can quantum-inspired primitives (e.g., lattice-like constructions) yield compact key update protocols suitable for fog devices?; How do these schemes compare to traditional key rotation in terms of bandwidth, latency, and resilience to node compromise?; How to integrate revocation efficiently across hierarchical fog clusters?
We will work by designing protocol primitives, proving basic security properties under standard assumptions, simulating hierarchical fog deployments, and implementing a proof-of-concept on Raspberry Pi clusters to measure performance.

24. 5G Network Slicing Controller for AR-assisted Industrial Maintenance with Digital Twins

We propose a network-slicing orchestration module that prioritizes AR telemetry and digital-twin synchronization for industrial maintenance tasks.
We ask: What slice parameters minimize AR latency while preserving other factory services?; How does dynamic digital-twin synchronization frequency affect operator performance and network load?; How to enforce QoS guarantees under fluctuating radio conditions?
We will work by building an emulated 5G testbed with slice management APIs, integrating an AR client and a lightweight digital twin, running user-in-the-loop maintenance scenarios, and measuring QoS, operator task time, and synchronization staleness.

25. Generative-AI Assisted Automated Bug Repair Pipeline for Legacy Enterprise Codebases

We propose a CI-integrated pipeline that uses generative models plus static analysis to propose, validate, and explain patches for legacy languages (e.g., COBOL, PL/I, RPG).
We ask: How effective are fine-tuned generative models at creating correct patches for legacy business logic?; How to combine static verification and test-suite augmentation to reduce false positives?; How to present patch rationales to maintainers to increase trust and adoption?
We will work by collecting legacy code corpora, fine-tuning models on repair tasks, integrating static analyzers and test generation, and running human-in-the-loop evaluations with enterprise developers on relevance, correctness, and trust metrics.

26. Privacy-Preserving Location-based Social Networking via Secure Multi-Party Computation and Differential Privacy

We propose a hybrid protocol combining MPC and differential privacy that enables proximity-based features without revealing exact locations to any server.
We ask: How can we achieve sub-50m proximity detection with acceptable latency under MPC constraints?; What differential-privacy parameters provide usable social features while bounding re-identification risk?; How to scale the hybrid protocol to thousands of simultaneous users?
We will work by designing the protocol, proving privacy guarantees, prototyping a mobile client and server-side MPC kernels, and evaluating latency, accuracy, and privacy leakage on emulated urban mobility traces.

27. Autonomous Ransomware Deception using Adaptive Honeypots Driven by Reinforcement Learning

We propose an RL-driven deception system that adapts honeypot behaviors and file-system artifacts to maximize attack engagement while gathering forensic intelligence.
We ask: Which RL reward structures lead to deceptive states that increase dwell time of ransomware actors without escalating risk?; How effective is adaptive deception at collecting novel payloads compared to static honeypots?; What safeguards prevent the system from unintentionally facilitating real breaches?
We will work by building a contained emulation environment, training RL agents to modify honeypot surface and telemetry, instrumenting ransomware families in sandboxes, and evaluating engagement metrics, data collection value, and ethical/safety constraints.

28. Low-code AutoML for Creating Accessible Web Applications for Users with Cognitive Impairments

We propose a low-code platform that auto-generates UI/UX variants tuned by AutoML to maximize comprehension and usability for people with cognitive impairments.
We ask: Which accessibility features (layout, language simplification, navigation aids) can be automatically selected per user profile to improve task success?; How do we measure and optimize cognitive load automatically from interaction signals?; How to integrate clinician feedback into the AutoML loop responsibly?
We will work by designing a low-code editor with variant generation, collecting interaction data from diverse user studies, applying AutoML optimization for accessibility metrics, and validating improvements through controlled usability experiments.

29. Blockchain-based Academic Credential Lifecycle with Verifiable Revocation and Privacy Controls

We propose a decentralized credential system that supports verifiable issuance, selective disclosure, and privacy-preserving revocation without centralized authorities.
We ask: How can we implement revocation that is cryptographically verifiable yet preserves graduate privacy?; What hybrid on-chain/off-chain patterns provide scalability and legal compliance?; How to design UX for institutions and employers that accepts cryptographic proofs with minimal friction?
We will work by designing smart-contract schemas, integrating privacy primitives (e.g., accumulators or succinct proofs), building issuer/verifier clients, and running pilot deployments with a university to measure latency, throughput, and adoption barriers.

30. Neuro-symbolic Policy Synthesis for Autonomous Drones in GPS-denied Indoor Environments

We propose a neuro-symbolic framework that fuses learned perception modules with symbolic planners to synthesize robust navigation policies for indoor drones without GPS.
We ask: How can symbolic constraints (e.g., no-fly zones, mission rules) be composed with learned controllers to guarantee safety?; How to perform sim-to-real transfer for perception modules while maintaining planner reliability?; How does the hybrid approach compare to end-to-end RL in sample efficiency and safety?
We will work by developing perception networks for feature extraction, encoding environment knowledge into symbolic planners, creating interfaces for runtime synthesis, testing in simulated indoor scenarios (ROS/Gazebo), and validating on real quadrotors in controlled indoor testbeds.

31. Federated continual learning for energy-harvesting IoT nodes

We ask: How can we design federated continual learning algorithms that adapt on intermittently powered, energy-harvesting IoT devices? We ask: How do we guarantee model convergence and bounded staleness under frequent power loss and partial participation? We ask: How do we minimize communication while preserving privacy and personalization for heterogeneous sensors?
We explain how to work: We design a simulator of energy-harvesting schedules and implement lightweight continual learning updates (e.g., reservoir sampling + elastic weight consolidation) on microcontroller-class devices. We collect representative sensor streams and run federated rounds under realistic duty cycles. We measure accuracy, communication cost, model drift, and privacy leakage. We prototype checkpoint/resume strategies and publish code and datasets for reproducibility.

32. Privacy-aware collaborative augmented reality (AR) annotation sharing

We ask: How can we enable ad-hoc, privacy-preserving sharing of AR annotations among users in public spaces? We ask: How do we encode spatial annotations so that access is limited by contextual predicates (time, co-location, social trust) without exposing raw scene imagery? We ask: How do we ensure low-latency synchronization for multi-user AR while enforcing privacy policies locally?
We explain how to work: We design a hybrid on-device/edge system that stores encrypted spatial anchors and policy tokens; we implement attribute-based encryption and searchable encryption for spatial queries. We run user studies with AR prototypes to measure usability and latency. We evaluate privacy leakage under threat models and iterate on UI affordances for consent.

33. Quantum-resistant over-the-air firmware updates for heterogeneous edge fleets

We ask: How can we build an over-the-air (OTA) update framework that uses lattice-based signatures and compact proofs to authenticate firmware across constrained devices? We ask: How do we manage key distribution, revocation, and rollback safety in large, heterogeneous fleets? We ask: How do we measure performance trade-offs of post-quantum primitives on real embedded hardware?
We explain how to work: We implement an OTA pipeline using NTRU/LMS-style signatures combined with Merkle proofs for chunked firmware delivery. We port primitives to representative MCUs and measure CPU, memory, and energy overhead. We design revocation protocols compatible with intermittent connectivity and test on a mixed-device lab fleet.

34. Explainable reinforcement learning for adaptive network QoS orchestration

We ask: How can we train RL agents to manage network resource allocation while producing human-interpretable explanations for actions? We ask: What explanation primitives (counterfactuals, feature-attributions, causal rules) are most useful to network operators? We ask: How do we maintain performance while constraining policy complexity for explainability?
We explain how to work: We build a network simulator with realistic traffic and let policy networks optimize throughput/latency under multi-objective rewards. We augment training with interpretable policy distillation and local explanation modules. We perform operator-in-the-loop evaluations to measure trust and debugging efficiency.

35. Decentralized edge identity management using sharded DAG ledgers

We ask: How can we design a lightweight, sharded DAG-based ledger for decentralized identity and attestation across edge devices with intermittent connectivity? We ask: How do we ensure fast local verification, cross-shard attestations, and privacy-preserving selective disclosure? We ask: How do we prevent sybil attacks in low-resource environments?
We explain how to work: We prototype a sharded DAG ledger with compact Merkle proofs and SPV-like verification for edge nodes. We integrate identity attestations, verifiable credentials, and reputation scoring with resource-based sybil resistance (e.g., hardware attestations). We evaluate latency, storage overhead, and resilience under churn.

36. Sensorless fault detection via ambient thermal and visual side-channels

We ask: Can we detect faults in electromechanical equipment using ambient thermal and passive visual side-channel patterns without instrumenting the equipment? We ask: Which features extracted from low-resolution thermal/video streams correlate reliably with common fault modes? We ask: How do we build models that generalize across device models and environments?
We explain how to work: We collect synchronized thermal and video recordings of devices under normal and fault conditions in controlled and field settings. We engineer spatio-temporal features and train cross-modal classifiers with domain-adaptation techniques. We validate detection latency and false-positive rates on unseen devices.

37. Zero-UI continuous authentication using micro-movement signatures and ultra-low-power sensors

We ask: How can we create continuous, zero-interaction authentication on personal devices using micro-gestural signatures from IMUs, touch capacitance, and acoustic sensors? We ask: How do we maintain low power consumption while keeping false accept/reject rates acceptable? We ask: How resilient are such signatures to mimicry and adversarial examples?
We explain how to work: We design a multi-sensor pipeline that extracts micro-movement features and trains lightweight authentication models using few-shot enrollment. We implement energy-aware sampling schedules and on-device anomaly detection. We run adversarial robustness tests including mimicry attempts and quantify energy/accuracy trade-offs.

38. Algorithmic auditing assistant for bias remediation in automated code review tools

We ask: How can we build an auditing assistant that detects and explains potential demographic or functional biases introduced by automated code review and CI/CD bots? We ask: How do we surface risky patterns without overwhelming developers with false positives? We ask: How can the assistant propose targeted remediation and measure downstream impact?
We explain how to work: We collect outputs from multiple automated code-review tools across diverse codebases and tag problematic review behaviors. We build explainable models that map tool outputs to risk categories and propose minimal-change remediations. We deploy in pilot CI pipelines and measure developer acceptance and bias reduction metrics.

39. Cross-modal malware detection using acoustic, electromagnetic, and power side-channels

We ask: Can we detect stealthy malware by correlating anomalies across acoustic emissions, electromagnetic signatures, and power consumption at the device level? We ask: How do we fuse heterogeneous side-channel streams to improve detection recall while controlling false alarms? We ask: How portable are learned detectors across hardware revisions?
We explain how to work: We instrument endpoints with passive acoustic, EM, and power sensors and collect baseline and infected behaviors using controlled malware artifacts. We design multi-stream fusion architectures and domain-adaptation layers to handle hardware variance. We evaluate detection latency, robustness, and deployability constraints.

40. Carbon-aware serverless task scheduling using reinforcement learning and forecasted grid mix

We ask: How can we schedule serverless function placement and timing to minimize carbon intensity without violating SLAs, using short-term forecasts of grid mix and renewable availability? We ask: How do we quantify trade-offs between carbon reduction, cost, and latency? We ask: How do we integrate forecasts and uncertainty into online scheduling decisions?
We explain how to work: We collect historical carbon intensity and serverless workload traces, and train RL schedulers that incorporate probabilistic forecasts. We simulate placement across multi-region clouds with cost and latency models and evaluate carbon savings under varying forecast accuracy. We prototype a scheduler plugin for an open serverless platform and measure real-world impact.

41. Edge-aware Federated Reinforcement Learning for Personalized UI Adaptation

We propose a system to personalize user interfaces on mobile/edge devices using federated reinforcement learning; we ask: How can we balance personalization and model convergence with heterogeneous edge resources? We ask: How does privacy-preserving reward sharing affect policy quality? We ask: What communication schedules minimize latency while preserving UX? We will collect simulated user-interaction traces and instrument a mobile testbed, implement federated RL variants (PPO/DQN) with differential privacy and adaptive client selection, and measure personalization gain, communication cost, convergence, and perceived UX in A/B studies.

42. Acoustic Ear-Canal Biometrics for Continuous Authentication on Earbuds

We investigate continuous authentication using in-ear acoustic responses captured by consumer earbuds; we ask: How distinctive and stable are ear-canal impulse responses across populations and time? We ask: Can lightweight on-device models detect impostors with low false-accept rates under movement and noise? We will record a private dataset across activities, design signal-processing pipelines to extract ear impulse features, train compact classifiers and anomaly detectors on-device, and evaluate robustness to noise, insertion variability, and replay attacks.

43. Quantum-Resistant Sharding Protocol for Blockchain-powered IoT Networks

We design a sharded blockchain protocol tailored to large-scale IoT with post-quantum signatures; we ask: Can sharding plus lattice-based cryptography meet latency and energy constraints of IoT gateways? We ask: How to dynamically re-shard under churn while preventing cross-shard attacks? We will model IoT network topologies, implement a prototype using lattice signatures (e.g., Kyber) and sharding strategies, simulate adversarial churn and quantum-capable adversaries, and measure throughput, latency, energy, and security margins.

44. Explainable Synthetic Data Generation for Fairness-aware Model Testing

We develop methods that generate synthetic datasets with explainable provenance to stress-test fairness properties; we ask: How to encode causal interventions in synthetic data so testers understand bias sources? We ask: Can explanation-guided generation improve discovery of fairness violations compared to random augmentation? We will formalize provenance metadata for generative models, build a generative pipeline (GAN/flow) conditioned on explainable interventions, run controlled experiments on benchmark fairness problems, and evaluate detection power and interpretability for practitioners.

45. Dynamic SDN-based Control for Platooned Autonomous Vehicles under Network Uncertainty

We propose an SDN controller that dynamically adapts control-message scheduling for vehicle platoons facing wireless uncertainty; we ask: How to co-design network control and vehicle controllers to maintain safety and throughput? We ask: What trade-offs exist between centralized SDN decisions and local failover policies? We will build a co-simulation of vehicular dynamics and SDN-enabled wireless, implement adaptive scheduling and local fallback controllers, run safety and throughput scenarios, and quantify platoon stability, control overhead, and resilience.

46. Energy-aware Neural Compilation for Microservice Function Chains on Heterogeneous Servers

We investigate compiler-level optimizations that map neural-network-backed microservices across CPUs, GPUs, NPUs to minimize energy under SLA constraints; we ask: How to compile service graphs to heterogeneous hardware for end-to-end energy reduction? We ask: Can dynamic recompilation using runtime telemetry outperform static placement? We will formalize an optimization problem, implement a prototype neural-aware compiler that uses profiling and cost models, test on representative microservice chains, and measure energy, latency SLAs, and overhead.

47. Zero-shot Malware Mutation Detection via Code-Behavior Embedding Alignment

We explore zero-shot detection of novel malware families by aligning static code embeddings with dynamic behavior embeddings; we ask: Can cross-modal alignment detect unseen mutation patterns without labeled examples? We ask: What embedding spaces and alignment losses best separate benign/malicious behavior? We will construct paired static/dynamic traces from benign and known-malware corpora, train alignment models (contrastive, mutual information), evaluate zero-shot detection on held-out mutated families, and measure detection rate, false positives, and interpretability.

48. Predictive Rendering for Holographic Telepresence to Compensate Network Latency

We design predictive-rendering pipelines to hide network latency in multi-user holographic telepresence; we ask: How accurate must motion and gaze prediction be to maintain presence without perceptual artifacts? We ask: How to synchronize cross-user predictions to avoid conflicting render states? We will collect motion/gaze datasets in telepresence tasks, develop lightweight temporal predictors and hybrid interpolation strategies, integrate into a prototype rendering pipeline, and run user studies measuring presence, motion artifacts, and perceived latency.

49. Real-time Carbon-aware Container Orchestration using Short-term Energy Forecasts

We build a scheduler that places containers based on short-term (hourly) renewable energy forecasts to minimize carbon impact while meeting performance; we ask: How to trade performance and carbon across geo-distributed data centers with forecast uncertainty? We ask: What forecasting horizon and uncertainty modeling yields best carbon savings? We will integrate renewable forecasts, uncertainty quantification, and a robust placement optimizer into Kubernetes, evaluate on traces and real cloud testbeds, and report carbon reduction, SLA adherence, and migration costs.

50. Privacy-preserving Graph Neural Networks for Federated Knowledge Graph Completion

We propose federated GNN training for distributed knowledge graphs with differential privacy and secure aggregation; we ask: How to preserve graph structural utility under privacy constraints across heterogeneous parties? We ask: Can partial graph alignment and compressed secret-sharing retain link-prediction accuracy? We will design federated GNN protocols with local subgraph sampling, DP noise budgeting, and cryptographic aggregation, run experiments on split KG benchmarks, and measure link-prediction performance, privacy loss, communication, and scalability.