╔══════════════════════════════════════════════════════════════════╗
║ ║
║ When a model is put under adversarial pressure — jamming, ║
║ evasion, obfuscation, injection — where does it structurally ║
║ break, and can that break be turned into a detector? ║
║ ║
║ Jailbreaks → swarm comms failure → biometric spoofing ║
║ → phishing infra hiding by renaming. ║
║ Same hunt. Four domains. ║
║ ║
╚══════════════════════════════════════════════════════════════════╝
| System | Domain | Headline | Status | Target |
|---|---|---|---|---|
| 🔴 EVELYN | Quantum graph ML | Name-invariant phishing infra detection via topology fingerprinting |  |
IEEE S&P / USENIX |
| 🟡 SWARN | MARL robustness | +13 pts detection under chase jammer · 1,050-run benchmark |  |
IEEE TNNLS |
| 🟢 MIMIC | Behavioral biometrics | JSD 0.1478 · 100% evasion on all detectors tested |  |
IEEE TIFS |
| ✅ ICEM | LLM red teaming | 🏆 Best Paper + Best Presentation |  |
NCTAAI 4.0 |
class AdversarialResearcher:
name = "Samratth Singh"
base = "DIAT · DRDO-affiliated · Pune, IN"
degree = "B.Tech CS (Cybersecurity)"
thesis = "Find the structural failure condition. Weaponize it. Then defend it."
domains = [
"Adversarial AI", # ICEM — jailbreak taxonomy
"LLM Red Teaming", # prompt injection / alignment evasion
"MARL Robustness", # SWARN — comms-degraded coordination
"Behavioral Biometrics", # MIMIC — mouse-dynamics forgery
"Quantum Graph ML", # EVELYN — topology-invariant detection
]
shipped = {
"Phish_Byte" : "F1 0.948 · 12K params · live on HuggingFace",
"T.E.M.P.E.S.T": "Windows attack-surface enumerator + ML anomaly scoring",
"RAV3N-sec" : "Static vuln scanner · 30+ vuln classes · 100% local",
}
proof = {
"best_paper" : "NCTAAI 4.0 · 2024",
"htb_rank" : "Top 5% Global",
"ctf" : "14th National · Ciphathon",
"ieee_targets": ["TNNLS", "TIFS", "S&P"],
}
def reach_me(self):
return ["research collabs", "red-team engagements", "MS opportunities"]🔴 EVELYN — Quantum Graph Phishing Infrastructure Detection
🔴 EVELYN — Quantum Graph Phishing Infrastructure Detection
They change their names. They can't change their shape.
Phishing campaigns cycle domains faster than any blocklist keeps up — but the infrastructure topology behind them stays constant. EVELYN runs a continuous-time quantum walk over a phishing-infrastructure hypergraph and extracts a topology fingerprint that is invariant to domain names — detecting a campaign by its shape, not its strings. Zero-shot on topologies never seen in training.
What I built:
- Reframed detection as a graph-spectral problem instead of string/GNN classification — the reason it generalizes zero-shot
- Robust to rapid domain rotation by construction, not by retraining
- Building empirical case for provable advantage over GNNs on infrastructure-shape tasks
Python · Continuous-Time Quantum Walks · NetworkX · DBSCAN · Graph ML
🟡 SWARN — Communication-Robust MARL for Adversarial Swarms
🟡 SWARN — Communication-Robust MARL for Adversarial Swarms
Most MARL assumes perfect comms. Real drone swarms never get that.
A 1,050-run benchmark exposing a structural blind spot: QMIX cannot perceive its own communication state, so under jamming it coordinates confidently on stale information. The fix: a two-process architecture — frozen QMIX explorer feeding a KL-anchored Navigator that stays stable when comms degrade.
┌─────────────────────────────────────────────────┐
│ Under Chase Jammer │
│ ┌──────────────────┬──────────────────────────┐ │
│ │ Baseline QMIX │ SWARN (Nav + QMIX) │ │
│ │ 42.5% │ 55.5% ▲ +13.0 pts │ │
│ └──────────────────┴──────────────────────────┘ │
└─────────────────────────────────────────────────┘
What I built:
- Diagnosed QMIX's blindness to comms state as structural, not a tuning issue
- Engineered frozen-explorer + KL-anchored Navigator, validated across full S1–S12 config grid
- 46 passing tests, VRAM guards, checkpoint auto-resume, strict env-parameter uniformity — reproducibility built in
PyTorch · QMIX · MAPPO · MARL · CUDA
🟢 MIMIC — Behavioral Evasion via Synthetic Mouse Dynamics
🟢 MIMIC — Behavioral Evasion via Synthetic Mouse Dynamics
The model that fooled every bot detector it met.
Hybrid LSTM + DDPM framework synthesizing human-indistinguishable mouse-motion trajectories — an adversarial stress test for behavioral-biometric defenses.
┌───────────────────────────────────────────────────────────────┐
│ Jensen–Shannon Divergence 0.1478 Near-human match │
│ Bot-detector evasion 100% Every detector beaten │
│ Custom behavioral dataset 212,568 Hand-collected corpus │
└───────────────────────────────────────────────────────────────┘
What I built:
- Combined LSTM temporal structure + DDPM stochastic realism — neither alone hits human-level JSD
- Built 212K-sample behavioral dataset from scratch
- Reframed as defense research: a generator this good is a benchmark forcing next-gen detectors to be robust
Python · PyTorch · DDPM · LSTM · Adversarial ML
9,000× smaller than DistilBERT. F1 0.948. From-scratch MLP on 29 hand-engineered features. Cascading inference: rule scorers in ~1ms, MLP in ~3ms. Per-feature attribution on every verdict.
|
Windows Attack-Surface Enumerator Read-only PowerShell recon — services, ports, autostart vectors, firewall rules, scheduled tasks, browser extensions. Optional Isolation Forest anomaly scoring. HTML dashboard.
|
Static Vulnerability Scanner AST + Regex hybrid catching 30+ vuln classes in Python. Severity-graded, rich CLI. 100% local — no API calls, no cloud, no telemetry.
|
Core
ML / Research
Offensive Security
2024 ─┬─ Aug git init — GitHub account created
├─ Oct SPH1NX · P.R.I.S.M · Phish_Byte shipped
└─ Dec 🏆 Best Paper + Best Presentation @ NCTAAI 4.0
2025 ─┬─ Jan T.E.M.P.E.S.T v1 — PowerShell + ML anomaly scoring
├─ Feb S.I.F.E.R · Sparakthon
├─ Mar Research Intern @ ICAR
├─ May Research Intern @ DIAT · DRDO-affiliated ← current
└─ Jun SWARN Phase A — 1,050-run benchmark complete
2026 ─┬─ Jan SWARN Phase D — Navigator+QMIX — +13 pts under jammer
├─ Feb EVELYN initiated — quantum graph phishing detection
├─ Mar MIMIC complete — JSD 0.1478 · 100% evasion
├─ Jun SWARN → IEEE TNNLS · EVELYN Stage 2 · Phish_Byte live on 🤗
└─ ?? Next structural failure condition...
┌──────────────────────────────────────────────────────────────────┐
│ │
│ Open to research collaborations, red-team engagements, │
│ CTF teams, and MS/PhD opportunities in adversarial AI. │
│ │
│ If you work where AI meets offensive security — let's talk. │
│ │
└──────────────────────────────────────────────────────────────────┘
// status: hunting structural failure conditions · last seen: DIAT, Pune