Raspberry Pi RL Security Defender

I’m building a C++ deep-reinforcement-learning module that runs natively on a Raspberry Pi 5 using TensorFlow Lite. I will stream JSON telemetry straight from the kernel—process activity, memory access patterns and network traffic—and the agent you write must learn, in real time, how to stop threats on its own. Detection and alerting are not enough; the objective is fully autonomous mitigation.

You’ll design and implement the RL logic (DQN, PPO or another on-policy/off-policy method you prefer) so that it compiles cleanly with CMake, links against TensorFlow Lite for ARM64, and can both train online and perform ultra-lightweight inference once deployed. The agent’s observation space is limited to the three signals above, and actions should map to practical counter-measures that can be executed from user space or via system calls.

Deliverables
• C++ source with clear build scripts for Raspberry Pi OS 64-bit
• Training loop, replay/experience buffer and checkpointing to SD-card or USB
• Inference module exposing a small CLI or REST endpoint for live mitigation
• Unit tests that feed mocked JSON sequences covering process spikes, memory anomalies and suspicious network flows
• README documenting setup, model architecture, and how to extend action handlers

Acceptance is met when I can flash a fresh Pi 5, run `cmake && make`, stream the sample JSON, see the agent converge, and watch it autonomously block or throttle the injected threats without human intervention.