20 AI Based FYP Ideas That Actually Impress Supervisors
Supervisors have seen plenty of attendance systems and sentiment analyzers. A project that tackles a problem people are worried about right now, like telling a real voice from a cloned one, tends to get a different reaction.
This guide gives you 20 AI FYP ideas you can actually finish. Most lean toward cybersecurity and deepfake detection, because those topics are current, have public datasets, and are easy to demo. Each idea has a difficulty level so you can match it to your skills and time.
What Makes an AI FYP Idea Strong?
Supervisors differ, but strong projects usually share the same traits:
A clear problem. You can explain who is helped in one sentence.
Real data. You use a public dataset or collect your own legally and ethically.
A baseline and an improvement. You compare your model with a simple one, not just report one number.
Honest evaluation. You go beyond accuracy and test on data your model has not seen.
A working demo. A small web app beats a notebook when you present.
Deepfake Detection Project Ideas.
A deepfake detection project works well for an FYP because the research community shares benchmark datasets, so you can compare your results with published work. Ethics note: use public research datasets and follow their licenses. Never create fakes of real people without consent.
# | Idea | What you build | Level |
|---|---|---|---|
1 | Face-swap video detector | Train a CNN or vision transformer on frames from FaceForensics++ and show which face regions drive each decision with Grad-CAM. | Intermediate |
2 | Voice-clone detector | Convert audio to spectrograms or use pretrained speech features to separate real from synthetic speech, using ASVspoof 5 or WaveFake. | Intermediate |
3 | Generalization benchmark | Train on one dataset, test on another (for example FaceForensics++ to Celeb-DF) and explain the performance drop. | Intermediate |
4 | Deepfake checker web app | Upload a short clip and get a real or fake score, a confidence value and a heatmap in a simple Streamlit or Flask app. | Intermediate |
5 | Lip sync mismatch detector | Compare mouth movement with the audio track and flag clips where the two do not match. | Advanced |
Idea 3 is a quiet favorite. The authors of the Celeb-DF dataset reported that it was significantly harder for existing detectors than earlier datasets, so a cross dataset study shows real research thinking.
If you like audio deepfake detection, ASVspoof 5 is the current challenge series. Its resources include around 2,000 speakers and attacks from 32 different algorithms, and they are freely available to the community.
Cybersecurity FYP Ideas With AI.
These cybersecurity FYP ideas suit students who want a mix of machine learning and practical security. They also read well on a CV for software house jobs.
# | Idea | What you build | Level |
|---|---|---|---|
6 | Explainable phishing detector | Classify emails or URLs from public phishing feeds and list the top reasons behind each verdict. | Beginner |
7 | Intrusion detection system | Train models on CICIDS2017 or UNSW-NB15 and compare random forest, XGBoost, and a neural network. | Intermediate |
8 | Malware classifier | Classify files as benign or malicious using static features from the EMBER dataset. You never run the samples. | Intermediate |
9 | Vulnerable code scanner | Use a fine-tuned or prompted model to flag risky code such as SQL injection, then compare it with a rule-based scanner. | Advanced |
10 | Log anomaly detector | Learn normal system behavior from logs and flag unusual events with an autoencoder or isolation forest. | Intermediate |
LLM Project Ideas for Students.
LLM projects are popular, so the trick is to pick one with a clear purpose and a way to measure success. These LLM project ideas double as AI security projects.
# | Idea | What you build | Level |
|---|---|---|---|
11 | Prompt injection guard | Build a filter that flags hidden or malicious instructions in text before it reaches an LLM app, and measure false alarms. | Intermediate |
12 | Tool description scanner | Scan AI tool and plugin descriptions for suspicious instructions and give each a risk score. | Intermediate |
13 | Cited document assistant | Create a retrieval-based chatbot over university or company documents that shows its sources and respects who can see what. | Intermediate |
14 | Misinformation checker | Retrieve evidence for a claim, then summarize whether sources support it, with links for every point. | Advanced |
15 | Security alert summarizer | Feed alerts to an LLM that groups, ranks, and explains them in plain language, with a human approving actions. | Advanced |
More Machine Learning Final Year Projects.
If you want a machine learning final year project with a simpler path, these ideas are easier to scope. Several have a local angle that international datasets often miss.
# | Idea | What you build | Level |
|---|---|---|---|
16 | Fraud detection system | Handle heavily imbalanced transaction data with resampling and judge it using precision, recall, and cost. | Intermediate |
17 | ID liveness check | Detect printed photos or replayed screens in a face login demo. | Advanced |
18 | Roman Urdu scam SMS detector | Classify scam and spam messages written in Urdu and Roman Urdu, a language mix English-only tools often handle poorly. | Beginner |
19 | Spoof proof attendance system | Build face recognition attendance that rejects photos and video replays. | Intermediate |
20 | Model trust dashboard | Add SHAP or Grad-CAM explanations, confidence scores, and error analysis to any project above. | Beginner |
Which Idea Fits You?
If you are… | Consider ideas. |
|---|---|
New to machine learning | 6, 18, 20. Small datasets, quick results, easy demos. |
Working on a normal laptop | 6, 10, 16, 18. No heavy GPU training needed. |
Aiming for a research style report | 2, 3, 5. Cross dataset testing gives you real findings to discuss. |
Aiming for a software house job | 7, 11, 13, 15. They show security thinking and product skills. |
Interested in both AI and web development | 4, 13, 20. Strong demos with a clear user interface. |
You can also combine two ideas. For example, pair idea 2 with idea 20, and you get a voice clone detector that explains why it flagged a clip. That one extra layer often makes a project feel finished.
Datasets to Start With.
Dataset | What it offers | Good for ideas |
|---|---|---|
FaceForensics++ | 1,000 original videos altered with four manipulation methods | 1, 3, 4 |
Celeb-DF | 590 real and 5,639 deepfake celebrity videos | 1, 3 |
DFDC | Large video set from the Facebook-led challenge on Kaggle | 1, 3, 4 |
ASVspoof 5 | Spoofed and genuine speech from about 2,000 speakers | 2 |
WaveFake | Synthetic audio from several generators across two languages | 2 |
CICIDS2017 / UNSW-NB15 | Labeled network traffic for attack detection | 7 |
EMBER | Static features for malware classification | 8 |
Always check each dataset’s license and download page before you start. Some require a request form.
How to Turn an Idea Into a Proposal
Write the problem in one sentence. Example: “Detect cloned voices in short audio clips.”
Pick your dataset and baseline. Decide what you will compare against.
Choose metrics early. Use precision, recall, F1, and AUC. For audio spoofing, equal error rate is common.
Plan the demo. Decide what a supervisor will see in five minutes.
Check compute needs. Video models train faster on a GPU, so plan cloud notebook time or use smaller models.
A Simple Project Timeline
Adjust this to your department’s calendar, but a typical shape looks like this:
Month 1: Read three or four papers, finalize the idea, and write the proposal.
Month 2: Download and clean the data, then train a simple baseline.
Months 3 and 4: Build and improve your main model, and test it on unseen data.
Month 5: Build the demo and add explanations or a dashboard.
Month 6: Write the report, rehearse the presentation, and prepare for questions on limits.
Starting the baseline early is the habit that saves most students. A weak baseline you finish beats a perfect model you never train.
Mistakes That Weaken an FYP.
Reporting one high accuracy score from a single dataset and calling it solved.
Copying a tutorial notebook with no original contribution.
Choosing an idea so big it cannot be finished in two semesters.
Ignoring ethics, licenses, and privacy when using faces, voices, or messages.
13. References
Li et al., Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics
Wang et al., ASVspoof 5: Design, Collection and Validation of Resources
Frank and Schönherr, WaveFake: A Data Set to Facilitate Audio Deepfake Detection
Deepfake Detection: A Comprehensive Survey from the Reliability Perspective
Frequently asked questions
What is a good AI FYP idea?
A good AI FYP idea solves a clear problem, uses real data, and can be finished in your time frame. Deepfake detection, phishing detection, and prompt injection filters are strong choices because they have public datasets and easy demos.
Is a deepfake detection project good for a final year project?
Yes. Public datasets such as FaceForensics++, Celeb-DF and ASVspoof 5 let you benchmark your work against published research. Start with one modality, video or audio, and keep the scope small.
Which dataset should I use for a deepfake detection project?
For video, FaceForensics++ is widely used and Celeb-DF is a harder test set. For audio, look at ASVspoof 5 or WaveFake. Check each license before downloading.
How do I start an audio deepfake detection project?
Pick a dataset such as ASVspoof 5, convert audio to spectrograms or use pretrained speech features, train a classifier, and report both accuracy and equal error rate. Then test on a different dataset.
What are good cybersecurity FYP ideas with AI?
Popular options include an explainable phishing detector, a network intrusion detection system, a malware classifier using static features, a log anomaly detector, and a code vulnerability scanner.
What are some LLM project ideas for students?
Try a prompt injection guard, a cited document assistant, a misinformation checker, or a security alert summarizer. Choose one with a measurable goal, such as detection rate or answer accuracy.
Are there AI project ideas for students with no GPU?
Yes. Phishing detection, scam SMS classification, fraud detection, and log anomaly detection run well on a normal laptop. Video deepfake models train much faster on a GPU, so plan cloud notebook time.
What is a good machine learning final year project for beginners?
A phishing or scam SMS classifier is a solid beginner choice. Add explanations and a simple web app to make it stand out.
How should I evaluate a deepfake detector?
Use precision, recall, F1 and AUC, and test on a dataset your model has not seen. A drop in performance across datasets is a normal finding and worth explaining.
What AI security projects can I build safely?
Focus on defense; detectors, scanners and classifiers. For malware work, use static features from a public dataset and never run live samples on your own machine.
How do I make my FYP stand out to supervisors?
Compare against a baseline, test on unseen data, explain your model’s decisions and show a working demo. A clear problem statement helps as well.
Can I use pretrained models in my FYP?
Usually yes, if you add your own work, such as fine-tuning, a new evaluation or a new application. Confirm your department’s rules with your supervisor.
Written by
Muhammad Sohaib