Drone Telemetry Anomaly Detection with Self‑Supervised JEPA

Detect tampered drone flights from telemetry alone—no anomaly labels required for training.

Course: TAMU CSCE 625 — Artificial Intelligence (Spring 2026) Team: Yaswanth Reddy Yaradoddi · Siva Sai Deepank Manoj · Ubaid Khan Mohammed GitHub: https://github.com/yaswanthreddyyyr/Drone-Anomaly-Detection-JEPA

TL;DR

Drone telemetry streams (GPS/altitude/speed/heading) can be spoofed or manipulated mid‑flight. We built a self‑supervised JEPA encoder that learns normal flight dynamics from normal logs only, then uses a Local Outlier Factor (LOF) detector on learned embeddings to flag suspicious chunks at inference time.

Best configuration (JEPA v3 + PCA + LOF‑Manhattan): AUC 0.7719, F1 0.685, composite Score 2.60 (+35.9% vs. our first JEPA setup).

1) Why this problem matters

Modern drone applications—delivery, agriculture, inspection, and public safety—depend on a continuous telemetry feed. If an attacker tampers with GPS or related signals (e.g., injecting fake waypoints, shifting timestamps, or causing coordinate jumps), the drone can deviate from its route, behave unpredictably, or crash.

The operational constraint is important:

• True attacks are rare (labels are expensive or unavailable).
• Detectors must be sensitive (catch real tampering) without producing constant false alarms.

This project targets that exact setting: anomaly detection without labeled attacks during training.

2) Data: DJI flight logs and anomaly types

We used DJI drone telemetry logs (Kaggle) with waypoint‑level binary labels for evaluation. Training uses normal flights only; anomalies are used only for validation/test evaluation.

Anomaly families (9)

Train/validation/test splits