ME Project Topics in Computer and IT Engineering

What separates an ME topic from a BE project

A BE project usually succeeds when the system works. An ME project succeeds when you can show that your method does something better, cheaper or more reliably than an existing method, under conditions you define and can repeat. That single difference changes how you pick a topic: start from a measurable weakness in published work, not from an application you would like to build.

Treat the topics below as starting points. Read recent papers and narrow the problem until it is specific to your own idea, because your examiner will ask why this gap is still open.

Security and privacy

1. Concept drift handling in network intrusion detection — Python, scikit-learn, River, public flow datasets
   Detection models lose accuracy as traffic patterns change. Design a drift detector that triggers selective retraining, and compare detection rate and retraining cost against periodic full retraining.
2. Measuring the cost of post-quantum TLS on constrained links — OpenSSL with ML-KEM support, liboqs, network emulation
   Larger keys and signatures affect handshakes on slow networks. Measure latency and packet behaviour for hybrid key exchange, then propose and test a tuning approach that reduces the overhead.
3. Lightweight phishing URL detection resistant to evasion — PyTorch, character CNNs, adversarial example generation
   Character-level models can be fooled by small URL changes. Build an adversarial training pipeline and evaluate how much robustness is gained for each unit of added inference time.
4. Log anomaly detection without fixed log templates — Transformers, Drain parser as baseline, public log datasets
   Template parsers break when software updates change log formats. Compare a template-free embedding approach with parser-based methods across simulated version changes.
5. Selective encryption for searchable cloud records — Searchable encryption libraries, PostgreSQL, leakage analysis
   Fully encrypted search is slow while weak schemes leak patterns. Propose a field-level policy that balances query speed against leakage, and quantify both on realistic query workloads.

Cloud, distributed systems and edge computing

1. Reducing serverless cold starts through invocation prediction — Time-series models, Python simulator, public function traces
   Keeping functions warm wastes memory, while cold starts hurt latency. Train a per-function arrival predictor and measure the trade-off against fixed keep-alive policies using trace-driven simulation.
2. Energy-aware pod scheduling in Kubernetes — Kubernetes scheduler framework, Go, Kepler, load generators
   Default schedulers do not consider power consumption. Write a custom scheduler plugin that uses estimated node power, then compare energy use and response time under mixed workloads.
3. Task offloading for edge devices under unstable connectivity — Python, deep Q-learning, iFogSim or a custom simulator
   Many offloading methods assume a steady link. Formulate offloading as a reinforcement learning problem with link failures included and test deadline misses against greedy heuristics.
4. Root cause localisation in microservices using traces — OpenTelemetry, Chaos Mesh, PyTorch Geometric, demo microservice apps
   A failure in one service shows up as symptoms in many. Model call graphs with a graph neural network and evaluate ranking accuracy under injected faults.
5. Adaptive compaction for LSM-tree storage engines — RocksDB, C++, YCSB benchmarks
   Fixed compaction settings suit only some workloads. Design a controller that changes strategy as the read-write mix shifts, and measure write amplification and read latency.

Networks and IoT

1. Learning-based congestion control fairness — ns-3, Mininet, reinforcement learning
   Learned congestion control can treat competing traditional flows unfairly. Add a fairness term to the reward function and evaluate throughput, delay and fairness index in mixed traffic.
2. Battery-aware LoRaWAN parameter assignment — ns-3 LoRaWAN module, Python analysis
   Adaptive data rate focuses on link quality and may drain some nodes early. Propose an assignment scheme that considers remaining battery, and measure network lifetime and packet delivery.
3. Compressing sensor streams for on-device anomaly detection — TinyML, TensorFlow Lite Micro, ESP32
   Microcontrollers cannot store much history. Study how much anomaly detection quality is lost at different compression levels and propose a compression method tuned to preserve anomalies.

Data engineering and software engineering

1. Learned cardinality estimation under shifting data — PostgreSQL, Python, Join Order Benchmark
   Learned estimators degrade when the data changes. Design an incremental update method and measure query plan quality and update cost against the built-in optimiser.
2. Automated data quality rules for streaming pipelines — Apache Kafka, Spark Structured Streaming, statistical profiling
   Handwritten validation rules miss new failure types. Infer constraints from historical data, then evaluate how many injected errors they catch and how many false alarms they raise.
3. Prioritising flaky test detection in CI pipelines — Python, gradient boosting, public flaky test datasets
   Rerunning tests to spot flakiness is expensive. Predict flaky tests from code and execution features, and measure CI minutes saved compared with rerun-only strategies.
4. Validating LLM-generated code with property-based tests — Hypothesis, open-weight code models, standard coding benchmarks
   Generated code can pass simple examples and still be wrong. Automatically derive property tests from specifications and evaluate how many extra defects they reveal.
5. Vulnerable dependency reachability analysis — Static analysis, Soot or CodeQL, public advisory databases
   Scanners flag vulnerable libraries even when the affected code is never called. Combine call graph analysis with advisory data, then measure the reduction in false alerts on open-source projects.

Applied machine learning for computing problems

1. Workload forecasting for cloud autoscaling — Temporal models, Kubernetes HPA, public cluster traces
   Reactive autoscaling responds after demand rises. Compare forecasting models by their effect on SLA violations and resource cost, rather than by forecast error alone.
2. Duplicate bug report detection across languages — Sentence transformers, FAISS, GitHub issue data
   Global projects receive reports in mixed languages, including Hindi and English. Evaluate multilingual embeddings and propose a retrieval method that handles code-mixed text.

How to test a topic before committing to it

Before you submit a synopsis, spend a week or two on these checks.

• Can you name a baseline? If there is no existing method to compare with, you have no way to show improvement.
• Is the data available today? Confirm you can download and open the dataset, or generate a realistic workload, before you depend on it.
• Can you reproduce one paper? Run or reimplement one closely related method. If you cannot, the topic may be too far beyond your current skills or hardware.
• Can you explain the gap in two sentences? If you cannot, read more before finalising the title.

Notes for SPPU and other university students

Students from Savitribai Phule Pune University and similar universities often ask whether a particular topic is allowed. That decision rests with your department and guide, so read the current rules issued by your own university and do not rely on a senior's old copy.

Road2Tech in Pune mentors master's students from topic selection through to implementation. You still do the research and need to understand each step, since you will be the one defending it.