Projects

FaithAudit: Do explainable deepfake detectors cite the region that caused their verdict?

2026 – present

Explainable forgery detectors are scored for plausibility or localization overlap, never for whether the cited region drives the verdict. I test that with counterfactuals that contain no synthetic pixels: every face-swap frame has a frame-aligned authentic twin, so the authentic region spliced back in is the counterfactual.

An identity splice controls for the seam. The codebase is complete and self-tested, with control detectors of known faithfulness, Qwen2.5-VL and CNN+Grad-CAM detectors, and a pre-registered development/test split. First runs on FaceForensics++ are scheduled.

PyTorchPython

Code

PredictorProbe: Is the I-JEPA predictor spatially grounded?

2026 – present

JEPA models train a predictor to infer a hidden region's representation from context and a positional query, then discard it. I query the frozen predictor at a patch position and rank its output against every patch's true representation, which gives localization accuracy with no labels or training.

Two chance baselines, a positional-query ablation, and cross-image controls bound the result. I designed the protocol and built the codebase; the first real-weights run is complete. Undergraduates I supervise run the experiments and the web instrument as their Final Year Design Project.

PyTorchPython

Code

Multiclass SQL Injection Detection with a 1D CNN–TinyBERT Model

2024 – 2025

I built a hybrid 1D CNN–TinyBERT classifier for multiclass SQL-injection payloads. Because payload classes are heavily imbalanced, I evaluated it with macro-F1 and PR-AUC under realistic adversarial scenarios, rather than accuracy alone.

PyTorchPython

Code

Lightweight LLM and Reinforcement Learning for Text-to-SQL Generation

2023

I built a TinyBERT-based text-to-SQL system on the Spider dataset. Adding reinforcement learning from execution feedback improved logical correctness without requiring a larger model.

PyTorchPython

Code