What is awesome-moai-security?

This repository is a curated reading list rather than a piece of software. It collects academic research papers about the security of mobile on-device AI, meaning AI models that run directly on a phone instead of on a remote server, using frameworks like LiteRT, Core ML, ExecuTorch, and ONNX. The list is a companion resource to a research survey paper written by the repository's authors, covering the attacks that threaten these on-device AI systems and the defenses researchers have proposed against them. The README organizes the field into a clear taxonomy. On the attack side, it groups papers into categories such as adversarial attacks that fool a model into wrong predictions, backdoor attacks that plant hidden malicious behavior into a model, attacks that steal a copy of the model itself, and attacks that drain a device's battery or slow it down. On the defense side, it covers techniques for hiding a model's internal details from attackers, verifying that only an authorized model can run, using secure hardware regions called trusted execution environments, and watermarking a model to prove ownership if it is stolen. For newcomers to this specific research area, the README includes a suggested reading order, starting with background studies on how common on-device AI already is in real apps, then moving into the core security risk of models being extracted or copied, followed by the specific attack surfaces, and finally the corresponding defenses. A table summarizes which security concerns fall under which broad category of protection, and the repository also lists open research problems and emerging directions for future work in this area. This is intended for security researchers, graduate students, and engineers who want a structured entry point into an academic field, not a tool you install or run. Its main value is as an organized bibliography with links to the original papers and, where available, their published code.