Security Researcher · Adversarial AI · Red Teamer
Pune · B.Tech Computer Science (Cybersecurity)
I research adversarial AI and LLM security — specifically how language models fail under adversarial pressure, and how to build systems that don't. On the applied side I build offensive and defensive security tooling: attack-surface enumerators, behavioral biometrics, phishing detection, and multi-agent tactical simulations.
Published researcher (Best Paper · NCTAAI 4.0). Currently working on communication-aware MARL under degraded real-world constraints — studying how RF-denied and high-dropout environments break swarm coordination, and building training regimes that are robust to it.
Privacy is a myth, but I'll try to exploit it anyway.
| Paper | Venue | Year | Status |
|---|---|---|---|
| ICEM: A Taxonomy Framework for LLM Jailbreaking and Prompt Injection | NCTAAI 4.0 | 2024 | ✅ Published · 🏆 Best Paper Award |
| MIMIC: Hybrid LSTM + DDPM Framework for Realistic Mouse Motion Synthesis and Behavioral Evasion | — | 2025 | ✅ Completed |
| Communication-Robust MARL for Adversarial Swarm Environments | Target: IJRR / Pattern Recognition | 2025–26 | 🔬 Active |
Most MARL research assumes perfect communication. Real drone swarms operate in RF-denied environments — links drop, latency spikes, topology shifts asymmetrically. This work asks: how does degraded communication change which algorithm wins, and can a policy be trained to be robust to it?
What's being built:
- A systematic comm-degradation benchmark across a dropout spectrum (0% → 80%), dynamic topology changes, and asymmetric link failures
- A comm-robust variant of MAPPO/QMIX explicitly trained under randomised dropout via communication curriculum
- The contribution is both the benchmark (quantifying degradation on swarm search) and the method (robust training regime)
Implemented inside ARKEN — a battlefield intelligence platform with live Bayesian threat inference, SRTM terrain modeling, and a Palantir Gotham-inspired ops UI. No LLM APIs. Pure probabilistic ML.
Why this matters: Sim-to-real transfer for drone swarms breaks primarily on communication, not control. A paper that quantifies the gap and proposes a training curriculum that survives it addresses an open problem that robotics, defence AI, and distributed systems communities all care about.
Target venues: International Journal of Robotics Research (robustness + real-world applicability) · Pattern Recognition (comm-robust features as representation learning)
Applying quantum walk-based graph encoding and hyperedge anomaly detection to classify phishing URL infrastructure at the graph topology level — moving beyond URL feature extraction into structural network analysis. Extends Phish_Byte.
| Project | What it does | Stack |
|---|---|---|
| T.E.M.P.E.S.T | Read-only Windows attack-surface enumerator · unsupervised ML anomaly detection · HTML dashboard output | PowerShell · Python |
| S.I.F.E.R | Behavioral biometrics system with iterative feedback · mouse dynamics · keystroke analysis | Python |
| Phish_Byte | Phishing URL detection · graph-based feature extraction · quantum walk research extension | Python · ML |
| SPH1NX | Network scan detector for TCP Null/UDP scans · JARVIS-style voice alerts · real-time dashboard | Python · Scapy |
| P.R.I.S.M | Port response identifier & service mapper | Python |
Languages
ML & Research
Security
Frameworks
Open to research collaborations, red-team engagements, and masters opportunities in adversarial AI and LLM security.
Reach me via LinkedIn