Adversarial Machine Learning
I build ML systems and study what happens when an adversary tries to break them. My work spans neural signal classifiers, surveillance recognition, and threat detection, with a focus on verifiable robustness under real-world conditions.
Final-year AI/ML student. I work at the intersection of machine learning and security — specifically on what breaks when a model faces an adversary, and how to build systems that hold up anyway.
My projects are benchmarks and tools, not demos. Each one produces a verifiable number on a real dataset — a result you can reproduce, critique, and build on.
Benchmark for adversarial robustness of EEG-based motor imagery classifiers. FGSM and PGD attacks tested across 9 subjects on BCI Competition IV-2a, with adversarial training and input smoothing evaluated as defenses. Inference latency measured against real-time prosthetic control constraints.
Two-layer cascading ensemble (LightGBM + TF-IDF N-gram LR) with SHAP per-prediction explainability. 124,902 URLs, 26 engineered features, served via FastAPI.
ArcFace-R50 fine-tuned on synthetically degraded LFW for real-world surveillance conditions — blur, low light, JPEG artifacts, sub-20px crops. SCRFD detection, GFPGAN restoration, FAISS vector index, judicial reproducibility protocol.
Production antivirus engine with four detection layers: Random Forest classifier, YARA signatures, heuristic analysis, and behavioural anomaly detection. Commercial product for VPN Digital Service Pvt. Ltd. Targets zero-day and polymorphic threats.
EEG classifiers in prosthetic control pipelines can be collapsed by imperceptible perturbations to the input signal.
BCIShield quantifies this across 9 subjects and two attack types, evaluates defense strategies, and identifies an unresolved gap between defended and clean-baseline accuracy — the open problem that motivates the next phase of work.
| Condition | Attack | Acc. | Δ |
|---|---|---|---|
| Clean baseline | — | 57.9% | — |
| No defense | FGSM ε=0.01 | 10.1% | −47.8pp |
| No defense | PGD ε=0.01 | 9.4% | −48.5pp |
| Input smoothing | FGSM ε=0.01 | 9.9% | negligible |
| Adv. training | FGSM ε=0.01 | 38.8% | +28.7pp |
| Latency | — | 5.98ms | <50ms |
Research projects, internships, or conversations about adversarial ML and AI security.
Open for research collaboration and internship opportunities.