AI-Driven Penetration Testing for ARM Systems: Experimental Evaluation and Deployment Framework Across Four Paradigms

ARM-based systems are widely deployed in edge and IoT settings, where penetration-testing tools must operate under strict memory, power, and thermal constraints, yet prior work offers limited deployment-oriented guidance for these platforms. We experimentally evaluated four AI paradigms—traditional machine learning (ML), deep learning (DL), large language models (LLMs), and reinforcement learning (RL)—across three ARM platform tiers (Raspberry Pi 5, Radxa ROCK 5B Plus, NVIDIA Jetson Nano), collecting over 8,000 individual metric observations—aggregating per-iteration latency samples, 100 ms-interval resource readings during sustained-load phases, and 24-hour stability measurements across all paradigms, configurations, and platforms—covering accuracy, memory, latency, power, and thermal behavior. Traditional ML achieved up to 98.1% accuracy with 162 MB peak memory usage, while a combined optimization pipeline for DL (pruning, INT8 quantization-aware training, distillation, and platform tuning) achieved 92.6% accuracy with a 93% model-size reduction. In contrast, the evaluated LLM configurations exhibited memory and latency requirements that exceeded practical edge budgets on these devices, whereas RL supported lightweight on-device inference via off-device training, using 52–254 MB during execution. We introduce the ARM AI Security Benchmark (AASB) framework to standardize evaluation methodology and provide deployment-oriented baselines for ARM platforms.