Computer Science Thesis Topics for BS, MS and PhD Students
Choosing a computer science thesis comes down to practical questions, and I want answers before I start reading papers.
- Can I download the data, or will I have to record, collect or label it myself, and how long will that take?
- Is my contribution a new model, a fairer evaluation, a working system or a study of how people actually use technology?
- Do I have the compute for my experiments, whether that is a laptop, a lab GPU or paid cloud credits?
- If I use data about people, such as voices, faces, messages or student records, what consent do I need?
- Will my baselines be strong enough that an examiner believes my improvement?
Computer science is too wide for one thesis, so a single list cannot cover it. Seven areas follow, chosen because Pakistani students can reach local data or local problems in each: Urdu language technology, computer vision, security and privacy, software engineering practice, networks and IoT, data science on public data, and trusted systems. Theory, formal methods, graphics and robotics hardware are left out, because a narrow topic in those areas needs its own verified sources.
Every topic shows a suggested level and the compute it needs. Final year BS students can shrink any topic to one component, one dataset and one strong baseline. MS students should add a comparison or an ablation, and PhD candidates need an original method or theory.
How to Choose a Computer Science Thesis Topic
A workable thesis states its contribution type, its data and its baseline before it names a model.
Decide what kind of contribution you are making
Examiners judge a new method, an empirical evaluation, a working system and a human study differently, so pick one. Testing existing models on new Pakistani data is a legitimate thesis. It is also often easier to finish than inventing a model.
Check data and compute first
Find the dataset, or a realistic plan to build it, before you read papers. Estimate the hours of labelling and the GPU time. If you need a GPU, confirm which lab server or cloud credit you will really have.
Choose baselines before models
Fix a simple baseline, such as TF-IDF with logistic regression or a threshold rule, and one strong published baseline. Report both. Without them, a high accuracy figure proves little.
Check consent, terms of use and licences
Voice, face, message and student data need consent and usually ethics approval. Public datasets and app stores also have terms of use. Common Voice, for example, forbids trying to identify speakers, so read the licence before you plan the experiment.
Urdu and Pakistani Language Technology Thesis Topics
Urdu is widely used but has far fewer public resources than English, so careful evaluation is itself a contribution. The National AI Policy 2025, approved by the federal cabinet in July 2025, includes support for a national language model with cultural and linguistic relevance.
Position correct as of September 2026; check the Ministry of Information Technology and Telecommunication for how the policy is being implemented.
1. Does Urdu Speech Recognition Work Equally Well for Speakers Outside Their Twenties?
How much does word error rate rise for older speakers and for accents missing from the training data, and can a small consented recording set close the gap?
The Urdu release 25.0 of Common Voice holds about 80 validated hours from roughly 500 speakers, and about nine in ten clips come from speakers in their twenties. Fine-tune a pretrained multilingual speech model, then report error rates by age band and accent on speech you record with consent. The licence forbids identifying speakers, so keep your recordings anonymous and separate from the training data. These figures come from the release 25.0 datasheet, so check the current release before you cite them.
Suits: MS | Compute: single GPU
2. Roman Urdu to Urdu Script Transliteration: Large Language Models Against a Rule-Based Baseline
Do large language models transliterate Roman Urdu more accurately than a hand-built rule system, and where do they fail?
Roman Urdu has no fixed spelling. Build a test set of 500 sentences with two annotators who write the correct Urdu script, then measure character error rate and the share of sentences that need no correction. Report agreement between the annotators, because their disagreement sets the ceiling for any model. Use the same test set for every system, and keep it out of any prompt examples.
Suits: BS final year, MS | Compute: laptop plus API credits
3. Do Large Language Models Answer Urdu Science Questions as Well as English Ones?
Is accuracy lower when the same question is asked in Urdu, and does the gap depend on the subject?
Write 150 school-level questions in English, translate each into Urdu, and have a second person check the translations. Then test two or three models on both versions with the same prompt format. Write your own questions, because published exam papers are copyrighted and may already sit in model training data. Report results by subject, not just as one overall score.
Suits: MS | Compute: API credits or single GPU
4. Retrieval-Augmented Answering Over Federal Statutes: Which Chunking Finds the Right Section?
Does splitting statutes by section retrieve the correct provision more often than fixed-length chunks?
The Pakistan Code is the Ministry of Law and Justice’s online repository of federal laws. Its own disclaimer says the content is for information and refers readers to the gazette notification for the original. Build a corpus from a small set of Acts and record the date you downloaded it. Then write 100 questions with a known correct section, and measure retrieval accuracy and whether the answer cites that section. Evaluate retrieval and citation only, because the system should never be presented as legal advice. For the legal side of the same statutes, see the law thesis topics collection.
Suits: MS | Compute: laptop plus API credits
5. Error Analysis of Open-Source OCR on Printed Nastaliq Urdu
Which kinds of material, such as newspaper columns, textbooks or photographed notices, give the highest character error rates, and which preprocessing steps help?
Urdu’s joined, slanting script makes OCR harder than it is for Latin text. Scan or photograph about 300 lines from three source types and type the correct text by hand. Then test an open-source engine such as Tesseract with and without preprocessing, and count errors by type: dots, joining forms and punctuation. Share only text you are allowed to share, because newspapers and books carry copyright.
Suits: BS final year, MS | Compute: laptop
6. How Unicode Letter Variants Split Urdu Search Results
How many search misses come from different Unicode forms of the same Urdu letter, and how much does normalisation recover?
Urdu text online mixes visually identical characters, such as different code points for yeh and kaf. Take a public news archive you are allowed to use, count the variants, and build a small set of queries with judged results. Then compare BM25 retrieval before and after normalisation. The project is compact and the result is easy to measure, so it suits a final-year timeline.
Suits: BS final year | Compute: laptop
Computer Vision Thesis Topics for Pakistani Settings
Vision projects fail more often because of data than because of models. Clinical image questions, such as chest X-ray or retinal screening, need hospital permission and are covered in the medical research topics collection.
7. Flood Extent Mapping From Sentinel-1 Radar in One Sindh District
Does a U-Net segment flood water more accurately than a simple threshold method, and does it hold up in a district it was not trained on?
Sentinel-1 radar images are free and see through cloud, which matters during monsoon floods. Use the 2022 floods, build labels from a published flood map or from Sentinel-2 images, and train on one district. Then test on another. Weak studies often split pixels randomly from the same scene, which leaks information, so split by area instead.
Suits: MS | Compute: single GPU
8. Cotton Leaf Curl Detection: Do Models Trained on Clean Images Work in the Field?
How much does accuracy fall when a model trained on tidy leaf photos is tested on field photos with soil, shadows and mixed backgrounds?
Cotton leaf curl disease is a long-standing problem for cotton growers in Pakistan. Work with an agricultural extension officer or a university crop lab to label images, and photograph plants in the field as well as against plain backgrounds. Split by plant and by field, not by image, or the model will memorise individual plants. Then report the drop between the two settings.
Suits: BS final year, MS | Compute: single GPU
9. Pakistan Sign Language Recognition That Works for a New Signer
How much does accuracy fall when the test signer was never seen in training?
Record signs from at least eight volunteers, with consent, in different rooms and lighting. A Deaf-community school or association can help you recruit volunteers and check that the signs are correct. Compare hand-landmark features with a classifier against a CNN on raw frames. Keep every signer’s frames in a single split, because mixing them inflates accuracy, and describe how you handled consent for video.
Suits: BS final year, MS | Compute: laptop
10. Motorcycle Helmet Detection at Night and With Pillion Riders in Karachi Traffic
How much does helmet detection accuracy fall at night, in rain and when a pillion rider partly hides the driver?
Record video from a fixed public vantage point with permission, label a few thousand frames, and test a pretrained detector before and after fine-tuning. Report results by condition, not as one average. Blur faces and number plates in anything you store or publish, and ask your ethics committee how to treat footage of members of the public.
Suits: MS | Compute: single GPU
11. Pothole Detection From Dashcam Video: Does Foreign Training Data Transfer to Local Roads?
How does a model trained on a public road-damage dataset perform on footage from local roads, and how much does a small local set add?
Check where your public dataset was collected. If it comes from another country, test whether the model transfers. Mount a phone on a vehicle, record routes in one city with GPS, and label a few hundred damaged frames. Then compare the foreign-trained model, a model fine-tuned on local frames and a mix of both. Hold out entire routes for testing, because neighbouring frames look almost identical.
Suits: BS final year, MS | Compute: single GPU
Cybersecurity and Privacy Thesis Topics
Security theses need clear ethics. Test only what you own or have written permission to test, and report findings in aggregate. Pakistan’s electronic crimes law covers unauthorised access, so ask your supervisor to approve the method before you begin.
12. Detecting Phishing SMS in Urdu, Roman Urdu and English
Do transformer models beat TF-IDF with logistic regression on a mixed-language message set, and how fast does accuracy fall as scam wording changes?
Collect messages donated by volunteers, remove names, numbers and links before storage, and have two people label them. Split by time, training on older messages and testing on newer ones, because scam wording changes. Report precision and recall for the scam class, since honest messages far outnumber scams.
Suits: MS | Compute: laptop or single GPU
13. A Security Checklist Audit of Pakistani Wallet and Banking Apps
Which controls from the OWASP Mobile Application Security Verification Standard are most often missing in publicly available local apps, judged by static analysis only?
Choose five to ten apps from public app stores, run static analysis tools on the downloaded packages, and score each against a fixed checklist. Report only aggregated, anonymised results. Do not publish exploit details, and follow responsible disclosure if you find a serious flaw. Confirm with your supervisor that your method stays within the law and within each app’s terms.
Suits: MS | Compute: laptop
14. Do Pakistani App Privacy Policies Meet the Principles of the Draft Data Protection Bill?
Which draft principles, such as consent, purpose limits and retention, do the privacy policies of popular apps address?
Sources dated to mid-2026 report that Pakistan had no enacted comprehensive data protection law. The Personal Data Protection Bill 2023 was approved by the federal cabinet but had not passed both houses of Parliament. Fix one version of the draft, turn its principles into a checklist, and have two coders score a sample of app privacy policies. Name the draft version in your title, because later drafts differ.
Position as reported in May 2026; check the Ministry of Information Technology and Telecommunication and the National Assembly’s list of bills for the current status.
Suits: BS final year, MS | Compute: laptop
15. Do Deepfake Audio Detectors Trained on English Catch Synthetic Urdu Speech?
How much does detection accuracy fall on Urdu speech generated by text-to-speech tools?
Generate synthetic Urdu clips with available text-to-speech tools, and record real speech only from volunteers who agree to take part. Test a published English-trained detector, then fine-tune it on part of your Urdu set. Evaluate on speech from a tool the detector never saw, because detectors often learn the quirks of one generator. Never clone anyone’s voice without their consent.
Suits: MS | Compute: single GPU
16. Cross-Dataset Generalisation of Intrusion Detection Models
Do models that score highly on one benchmark, such as CIC-IDS2017, still work on another, such as UNSW-NB15?
Train on one dataset and test on the other after mapping the features to a common set. Duplicate rows and labelling errors are known problems in these benchmarks, so check for them first. Then report the drop between within-dataset and cross-dataset results, and discuss why it happens.
Suits: BS final year, MS | Compute: laptop or single GPU
17. Password Reuse Across Banking and Social Apps Among University Staff
What beliefs and habits lead staff to reuse passwords, and does a short training session change stated intentions?
Use a survey followed by a few interviews. Never ask for real passwords, only for habits and reasons. A before-and-after design with a short training session gives you something measurable. State the limits of self-report plainly, because people tend to describe safer habits than they practise.
Suits: BS final year, MS | Compute: none (survey)
Software Engineering Thesis Topics From Pakistani Practice
Software engineering theses study how software is actually built and used. Company access is the hard part, so agree on confidentiality in writing before you ask for data.
18. How Accurate Are Sprint Estimates at One Software House?
How far do story-point estimates differ from actual effort over twelve months, and which kinds of task are most often underestimated?
Ask one company for anonymised issue-tracker exports, with names and client details removed, under a written confidentiality agreement. Then compute estimation error and compare task types, developer experience and sprint length. Tickets often lack logged hours, so check completeness before you promise any analysis.
Suits: MS | Compute: laptop
19. Usability of the HEC Online Degree Attestation Workflow
Can recent graduates complete an attestation application without errors, and where do they get stuck?
HEC introduced a fully online, paperless degree attestation system, effective 11 May 2026, with applications submitted through its e-services portal. A real application needs personal documents and a fee. So observe graduates who are applying anyway and agree to be observed, or run a heuristic evaluation of the public screens with trained evaluators. Measure completion time and errors, and score satisfaction with the System Usability Scale. Tell participants that your study is independent of HEC.
Position correct as of September 2026; check HEC’s website for how the workflow has changed.
Suits: BS final year, MS | Compute: none
20. Do AI Coding Assistants Help Junior Developers Finish Tasks Faster Without More Defects?
How do completion time and defect count change when graduates use an assistant on a standard task?
Run a controlled experiment with 20 to 30 final-year students or fresh graduates, using two matched tasks and a crossover design so each person works with and without the assistant. Measure time, passing tests and a code-quality score from two reviewers. Practice effects can hide the real difference, so randomise the task order. Name the tool and version you used.
Suits: MS | Compute: laptop
21. Startup Time and Memory Use of Popular Local Apps on Low-Memory Phones
How much slower do local apps start, and how much memory do they use, on a phone with little RAM compared with a mid-range phone?
Pick a fixed set of public apps and test them on two or three phones with different RAM. Measure cold start time and memory over repeated runs with the platform’s profiling tools. Report medians and spread, because single runs are noisy. Name the app versions too, since results change with each update.
Suits: BS final year, MS | Compute: test phones
Networks, Cloud and IoT Thesis Topics
Hardware and field experiments make strong theses because the data is yours. Budget time for borrowing equipment, and allow for power cuts and weather.
22. Calibrating Low-Cost PM2.5 Sensors in Lahore
Does a humidity-corrected model make low-cost sensor readings agree with a nearby reference monitor, and does the correction hold across seasons?
A 2024 note from the LUMS city lab described a few government reference-grade monitors, one US consulate monitor run through AirNow, and public low-cost sensors in Lahore. Check what exists today. Pair a low-cost sensor’s readings with the nearest reference monitor, then compare a simple linear correction with a random forest. Train on one season and test on another, and check the data owner’s terms before you download or share readings.
Suits: MS | Compute: sensors plus laptop
23. Duty-Cycling Soil-Moisture Nodes to Survive Power Cuts
How does the sampling interval trade off battery life against irrigation decision accuracy when the gateway loses mains power?
Build a small network of microcontroller nodes with soil sensors and a gateway with battery backup. Log how long the system runs through simulated and real outages, and compare fixed against adaptive sampling intervals. Estimate accuracy against readings taken at a much shorter interval during a test period. Field tests need landowner permission and time for weather.
Suits: BS final year, MS | Compute: microcontroller boards
24. Mobile Network Quality Across Three Islamabad Neighbourhoods
How do download speed and latency differ across operators, locations and times of day?
Write or adapt a measurement app that records speed, latency and signal strength with timestamps. Collect repeated samples at fixed spots and times from volunteers using different operators. Test differences statistically, not by eye. Strip precise locations from any published data, and describe your sampling plan honestly, because crowdsourced samples are not random.
Suits: BS final year, MS | Compute: phones
25. Cloud Rental or Lab GPU Server? A Cost Model From One Department’s Usage Logs
At what monthly usage does owning a GPU server cost less than renting equivalent cloud capacity?
Collect a semester of job logs from a lab with permission, anonymised by user. Build a cost model with purchase price, power, maintenance and cloud rates as variables, and show how the break-even point moves. Prices and exchange rates change, so present the model with your inputs and the date you took them.
Suits: MS | Compute: laptop
26. LoRa Range in Dense Urban Streets Versus Open Ground
How does packet delivery fall with distance in a dense city area compared with open terrain?
Use off-the-shelf LoRa modules, a fixed gateway and a moving transmitter with GPS. Log signal strength and packet loss at set distances in both settings. Check the Pakistan Telecommunication Authority’s rules for the frequency band and power you plan to use before you transmit.
Suits: BS final year, MS | Compute: LoRa modules
Data Science and Machine Learning Thesis Topics on Pakistani Data
Public data makes these topics easier to start, but check coverage, missing values and licence terms first. Business-facing forecasting questions, such as sales or inflation, are covered in the business administration thesis topics collection.
27. 24-Hour-Ahead PM2.5 Forecasting for Lahore in the Smog Season
Does adding wind and humidity to past pollution readings improve forecasts beyond a persistence baseline?
Use hourly readings from one reference station and weather variables from a public source. The persistence baseline, which predicts that tomorrow equals today, is hard to beat in stable weeks, so include it. Evaluate across a whole smog season with time-ordered splits, and report error separately for high-pollution days.
Suits: MS | Compute: laptop
28. District-Level Wheat Yield Prediction From Satellite Vegetation Indices
Do vegetation indices from freely available satellite images predict district wheat yield better than last year’s yield alone?
Obtain district crop statistics from the provincial crop reporting service where available, and check which years and districts they cover. Compute seasonal vegetation indices from open satellite imagery, then compare gradient boosting with a linear model. Use leave-one-year-out validation, since random splits leak weather patterns across years.
Suits: MS | Compute: laptop
29. At-Risk Student Prediction Without Unfair Error Rates
Can a model flag students at risk of failing a course early in the semester, and are its error rates similar across gender and programme?
Use one department’s anonymised records, with permission from the registrar or head of department. Use only information available early, such as attendance and first quiz marks, not final grades, or the model will look better than it is. Report false-negative rates by group, not only overall accuracy. Then discuss what a department should and should not do with a risk flag.
Suits: MS | Compute: laptop
30. Predicting T20 Match Outcomes Ball by Ball in the Pakistan Super League
How well does a model estimate win probability after each over, and does it beat a simple run-rate rule?
Build features such as runs required, wickets in hand and balls left, and compare logistic regression with gradient boosting. Check that the ball-by-ball records you use cover every season you need, and cite their source. Calibration matters more than raw accuracy here, so plot predicted win rates against actual ones.
Suits: BS final year, MS | Compute: laptop
Trusted Systems and Blockchain Thesis Topics
Blockchain topics are strongest when they ask whether a ledger is needed at all, instead of assuming that it is.
31. Verifiable Degree Credentials Compared With Online-Verified e-Attestation Certificates
Which properties, such as tamper evidence, revocation and privacy, does each approach give, and at what cost in complexity?
HEC’s revamped system issues e-attestation certificates that can be verified online, and press reports quote HEC’s chairman describing it as built on blockchain technology. You will not have access to its internals, so compare only publicly documented behaviour. Build a small prototype with the W3C verifiable credentials model, then compare both approaches against a written threat model.
Position correct as of September 2026; check HEC’s website for how the system is currently described.
Suits: MS, PhD | Compute: laptop
32. When Is a Blockchain Unnecessary? A Signed Database Against a Permissioned Ledger
For one workflow with a small set of known parties, does a ledger offer any measurable benefit over a database with signed, append-only logs?
Pick a workflow such as tracking the stages of a scholarship application or a donation. Build both designs and compare latency, throughput, auditability and operating effort. List the assumptions that would make the ledger worthwhile, for example parties who do not trust a single operator. A negative result is acceptable if your comparison is fair.
Suits: MS | Compute: laptop
Match Your Topic to the Evidence You Can Reach
| Evidence route | Topics | Confirm first |
|---|---|---|
| Public data or documents | 1, 4, 6, 7, 14, 16, 27, 28, 30 | The data covers your period, its licence allows your use, and you can rebuild your splits |
| Data you must record, label or measure | 2, 3, 5, 8, 9, 10, 11, 12, 13, 15, 21, 22, 23, 24, 26 | You have the hours, consent and equipment, and a clear labelling guide |
| Participants (surveys, observation, experiments) | 17, 19, 20 | Ethics approval, a recruitment plan and a realistic sample size |
| Records from one organisation, with permission | 18, 25, 29 | Written permission and anonymisation before you receive any data |
| Design and prototype | 31, 32 | A fair comparison design with a stated baseline |
Scaling a Topic to Your Level
BS final year. For computer science project topics, treat any idea above as a small system with a clear test: one component, one dataset and one strong baseline. A working, well-evaluated small system beats an ambitious unfinished one.
MS. Add a comparison, an ablation or a second dataset. Then state what your study adds, whether that is a new setting, a fairer evaluation or a better measure.
PhD. Your thesis needs an original method or theory, so topics such as 16 and 31 work only as starting points. For help planning chapters and methods around committee feedback, see PhD thesis planning support.
Crowded Computer Science Research Topics and How to Narrow Them
| Crowded idea | Why examiners push back | Narrower version |
|---|---|---|
| Intrusion detection on a single benchmark | High scores often reflect duplicates and leakage | Train on one dataset and test on another, as in topic 16 |
| A chatbot for a university | A demo without evaluation convinces no one | Collect real student questions, define tasks, and measure retrieval accuracy and failure cases |
| Blockchain for a records system | A database often does the same job | Compare against a signed database, as in topic 32 |
| Face-recognition attendance | Similar demos are common, and privacy and bias concerns follow | Measure false-match rates across lighting with consenting volunteers, and justify why faces are needed |
Your Next Step
Shortlist two or three topics and test each with three questions. Can you reach the data? Do you have a strong baseline and the compute to run it? Do you have the consent and permissions you need?
For several years, students have come to us with a stalled chapter, a supervisor’s red pen and a deadline that will not move. From first-year BS projects to PhD chapters, students have trusted us with the parts of their research they found hardest. If you want feedback on a shortlist or help shaping one topic into a proposal, send your details through the order form for a quote. Reference material is meant for learning, so use it according to your institution’s policy. You can also read what other students have written about the support.