EFFICIENT AND TRUSTWORTHY ARTIFICIAL INTELLIGENCE SYSTEMS: EDGE MODEL OPTIMIZATION AND ADVERSARIAL ROBUSTNESS INLARGE LANGUAGE MODELS

Abstract

Modern AI systems must satisfy two distinct but equally important requirements: efficiency, the ability to deliver intelligence under tight memory, latency, and energy budgets at the edge of the network; and trustworthiness, the ability of large generative models to remain factually reliable and robust against adversarial manipulation. This thesis investigates both through two complementary contributions. The first contribution is a structured survey of Edge AI covering 2015–2025. It analyses lightweight CNNs (MobileNet, EfficientNet, SqueezeNet, ShuffleNetV2), real-time object detectors (YOLOv8, SSD-MobileNet), TinyML systems (MCUNet, DeepConv-LSTM, TinyBERT), and sub-10B SLMs (LLaMA 3, Phi-3 Mini, Qwen2.5 VL, GLM-4), with four compression and routing techniques: quantization, pruning, knowledge distillation, and Mixture-of-Experts. Twelve representative models are com pared across accuracy, parameter count, memory, target hardware, and trade-offs. The survey is supplemented by a focused experimental validation of the INT8 quantization trade-off on MobileNetV2: three TFLite regimes (FP32, INT8 dynamic-range, INT8 full-integer) are compared on Imagenette, confirming the published ∼3.5× size com pression at ≤1.3 percentage points accuracy drop and revealing that the latency benefit of quantization is hardware-dependent. Open challenges include activation-memory bottlenecks, hardware portability, on-device training, multimodal edge intelligence, energy benchmarking, and edge security. The second contribution is a systematic survey of adversarial hallucination attacks against LLMs, comparing open-source models (LLaMA 2, Mistral 7B, Falcon 40B) and closed-source models (GPT-4, Gemini 1.5 Pro, Claude 3 Opus). It introduces a taxonomy by attacker access: white-box (gradient and fine-tuning backdoor attacks, open-source only), black-box (factual corruption, CoT hijacking, authority persona; both classes), and RAG-layer (knowledge-base poisoning and indirect injection; both classes). Benchmarks (TruthfulQA, HaluEval, FActScore) are reviewed for adversarial suitability and shown to be inadequate for measuring robustness under active attack. The central finding is that closed-source release does not buy safety against retrieval-layer at tacks: while open-source weights expose a wider attack surface (gradient-based attacks, fine-tuning backdoors), closed-source models—despite stronger alignment—remain v equally vulnerable to RAG injection because retrieval is architecturally decoupled from alignment. A risk-graded attack-surface matrix and an evidence-based rubric for the High/Medium/Low risk grading are provided. Together, the contributions give a unified view of contemporary AI deployment: edge optimization explains how intelligence can be made efficient enough to run close to the data, and adversarial robustness explains how it can be made trustworthy enough to rely upon. The thesis identifies the security of compressed and quantized models— the very models deployed at the edge—as the single most important cross-cutting research gap, supports the claim with a meta-analysis synthesised from prior work, and outlines future directions in hardware-algorithm co-design, secure RAG, adversarial hallucination benchmarking, and alignment-as-security.