Event language
UI language
<p>Edge AI is growing fast. Small devices now run small models that make real decisions in homes, factories, and cities. But these models face the same threats as large AI systems while running with far less compute, memory, and security tooling. This talk introduces a practical security playbook for protecting AI workloads on resource-constrained edge devices.</p><p>We will explore common attack paths like model extraction, input probing, data poisoning, adversarial triggers, and runtime manipulation. Then we look at how open-source tools can defend these devices without slowing them down. The talk covers lightweight model integrity checks, telemetry using eBPF-style tracing, secure update pipelines, sandboxing patterns, and methods for detecting unusual inference patterns on tiny models.</p><p>The goal is to show how developers can secure edge-AI systems with fast, transparent, and community-friendly techniques. The session ends with a working demo of a minimal edge AI security setup: a small model on a low-power board, traced and protected using open tooling.</p>
