Final-year B.Tech (CS, Cybersecurity) · Research intern at DIAT (DRDO-affiliated), Pune
I work on adversarial robustness of intelligent systems — finding where models structurally break under pressure (jamming, evasion, injection, obfuscation) and turning those failure modes into either detectors or defenses.
- 🌐 Portfolio
- 🤗 HuggingFace
Most of the code below lives in private repos while papers are under review. Happy to share details on request.
| Project | Area | Status |
|---|---|---|
| EVELYN | Quantum graph ML for phishing infrastructure detection (name-invariant, topology-based) | In progress · targeting IEEE S&P / USENIX |
| SWARN | 12-config MARL benchmark, 7 algorithms, 5–100 agents, zero-shot transfer | Under review · Swarm and Evolutionary Computation (Elsevier) |
| Navigator + QMIX | Two-process MARL for jamming-resilient coverage; frozen QMIX explorer + KL-anchored PPO navigator | Under review · IEEE TCCN |
| MIMIC | LSTM + DDPM synthetic mouse dynamics; JSD 0.1478, 100% evasion on tested detectors | Under review · IEEE TIFS |
| ICEM | Taxonomy of LLM adversarial attacks | Published · NCTAAI 4.0 (Best Paper + Best Presentation) |
Vanilla QMIX is comm-blind: a reactive jammer barely degrades it because the learned coordination was never conditioned on communication content. Freezing QMIX as an exploration prior and training a KL-anchored PPO Navigator on top recovers comm-aware behavior.
| Method | Clean | Dropout | Static jam | Reactive jam |
|---|---|---|---|---|
| QMIX | 40.5% | 41.5% | 42.0% | 42.5% |
| IC3Net | 39.5% | — | 49.0% | 46.0% |
| OW-QMIX (α=0.5) | 63.5% | — | 60.0% | 55.5% |
| Navigator + QMIX | 50.0% | 49.5% | 54.5% | 55.5% |
| Δ vs QMIX | +9.5 | +8.0 | +12.5 | +13.0 |
20 seeds/cell, paired t-tests, 46 passing tests.
Phishing email classifier. 250K-param MLP over 85 hand-engineered features, cascading inference (rule scorers ~1ms, MLP ~3ms), per-feature attribution on every verdict. F1 0.950, ~1,500 emails/s. 130+ downloads.
Read-only Windows attack-surface enumerator in PowerShell. Services, ports, autostart, firewall, scheduled tasks, browser extensions. Optional Isolation Forest anomaly scoring, HTML report.
Static Python vulnerability scanner. AST + regex hybrid covering 30+ vuln classes, severity-graded, 100% local (no telemetry).
Languages: Python, C, TypeScript, PowerShell ML: PyTorch, scikit-learn, NumPy, Pandas, HuggingFace Transformers, CUDA, NetworkX Web / systems: FastAPI, React, Three.js, MediaPipe Security: Burp Suite, Metasploit, Nmap, Wireshark, OWASP ZAP, Scapy, Ghidra, SQLMap, Volatility, Aircrack-ng
- HackTheBox — top 5% global
- Ciphathon — 14th nationally
- Chair, ACM Student Chapter, DYPIU
Open to research collaborations, red-team engagements, and MS/PhD conversations in adversarial ML and AI security.