Suyash Singh

A drone with a swapped model, or a camera fooled by a printed patch, keeps sending actions that look perfectly normal. VeriSwarm makes the rest of the swarm catch it.

After every inference, a drone signs a receipt that binds who it is, the frame it saw, the model it ran and the action it chose — under one signature made inside an OP-TEE secure element. Peers verify the signature, check the model hash against an approved list, and only where their camera footprints overlap, compare the claimed action with what they saw themselves.

Disputes go to a reputation-weighted Byzantine quorum that tolerates ⌊(k−1)/3⌋ lying voters. A drone that keeps losing consensus decays to a floor weight and is dropped. The protocol logic is pinned down by 102 property-based tests.

First author · Second author: Dr. Subbulakshmi T, VIT Chennai

ReceiptEd25519

R = (d, t, hₓ, m, a, ν)
σ = Sign_sk(Canonical(R))

d: drone identity
t: timestamp, ns
hₓ: SHA-256 of the frame
m: SHA-256 of the model
a: chosen action
ν: fresh 16-byte nonce

6.2–19.3 ms

Live consensus, 3 → 11 drones, real LAN

0 / 4,000

False rejections on honest receipts · 95.5% without the co-visibility gate

4.7 ms

Hardware-signed receipt, 21 mJ each

1.5–7.6 KB

Attestation traffic per cycle

Measured on the Jetson Orin Nano — the slowest device in the swarm, so these are worst-case numbers.

VeriSwarm node architecture: input block, four protocol phases, and applied outcome. — **Architecture**One inference cycle: input, the four protocol phases, and what happens next — execute, isolate, or defer.

Adversarial patch transfer across viewpoints. — **Patch transfer**A patch built for one camera blinds its victim (0.88 → 0.00 confidence) but stops working beyond ~9° of viewpoint. At 23° the neighbour re-detects the obstacle in every trial.

ROC sweep used to choose the agreement threshold theta. — **Choosing θ**The agreement threshold sits at the bend of the sweep — detection near its peak, false positives still zero, stable across ten seeds.

The honest part: the semantic check has a limit. A patch deliberately optimised to work across viewpoints does defeat it — so the paper measures that case instead of citing its way around it.

Built, not only simulated

The hardware root of trust is mine end to end: a custom OP-TEE Trusted Application that signs inside the ARM TrustZone secure world on a physical Jetson Orin Nano. I built the airframe around it and ran the system both in PX4/Gazebo simulation and in real flight.

Airframe

Python · gRPC / protobuf · PyTorch · YOLOv8n · OP-TEE · ARM TrustZone · Ed25519 · PX4 SITL · Gazebo · Jetson Orin Nano

Manuscript in preparation for an IEEE journal.

Most messengers pick a side. GridCC refuses to.

Signal picks secrecy and makes multi-device painful; Instagram picks convenience and reads everything. GridCC runs a Signal-Protocol Double Ratchet on the wire — forward secrecy and post-compromise security — with a layered key-recovery scheme, so signing in on a new device restores full history in seconds with no QR scanning.

X25519 identity keys per device, AES-GCM-256 envelopes, HKDF-SHA256 deriving every chain and message key, private keys held in the hardware-backed Keystore on Android and the Keychain on iOS. Built twice, natively — Kotlin with Jetpack Compose, and Swift with SwiftUI — over a Node, MongoDB and Socket.io backend.

Built with Abhijan Patra

Kotlin · Jetpack Compose · Swift · SwiftUI · X25519 · Double Ratchet · AES-GCM-256 · HKDF · Node · Socket.io · MongoDB · Firebase

The thing that locks your machine during an exam is, itself, an attack surface.

Neo Browser locks a machine down during proctored academic assessments, which makes the strength of that lockdown the whole product. I tested it, found vulnerabilities, and scoped them together with the IamNeo.ai team and VIT Chennai faculty — working with the vendor rather than around them.

Everything went back through a structured responsible-disclosure report. The finding is rarely the interesting part of security work; doing something useful with it is.

Application security · Vulnerability research · Responsible disclosure · Vendor coordination

A zero-trust control plane that fits in your pocket.

A Raspberry Pi Zero W runs headless Kali Linux with no data link to the phone at all — every byte crosses the phone's own Wi-Fi hotspot. Authentication is strictly Ed25519 SSH keys held in the Android Keystore, with TOFU host-key pinning so the first connection can't be silently intercepted.

PHONE ── hotspot ──▶ PI ZERO W :22 ── ssh -L ──▶ VNC @ localhost:5901

VNC and websockify stay bound to localhost and are reached only through SSH local port forwarding; UFW limits the Pi to port 22 from the hotspot subnet. Discovery runs over mDNS with a parallel TCP-22 sweep as a fallback, and a foreground service keeps tunnels alive through long jobs.

Kotlin · Jetpack Compose · Ed25519 · TOFU pinning · SSH tunnelling · Kali Linux · Raspberry Pi · mDNS / Avahi · UFW

Suyash
Singh

1Sign

2Verify

3Vote

4Isolate

Work

Built, not only simulated

Writing

Roles

Field notes