Deep neural networks (DNNs) have achieved notable success in mobile vision tasks, yet they show vulnerability to adversarial attacks. When carefully crafted perturbations are introduced, these models can be easily misled into wrong classifications, posing significant risks for safety-critical mobile systems like autonomous vehicles. Although various defense strategies, both static and dynamic, have been proposed, many fail to address adaptive attacks or overlook the resource constraints of mobile systems. To address these limitations, in this paper, we present GuSoDrop, a lightweight dynamic defense framework that applies stochastic layer dropping. GuSoDrop leverages randomness to counteract adaptive attacks while selectively dropping less important layers to reduce computation overhead. Our preliminary evaluation shows that GuSoDrop outperforms state-of-the-art defense methods against different adaptive attacks and improves efficiency in reducing computational overhead.

Deep neural networks have advanced the perception and decision-making functions of smart embedded systems, such as car-borne driver assistance. Deploying these embedded neural networks often faces two challenges: (i) security vulnerabilities to adversarial examples that can be deployed in the perceived physical environment; (ii) limited computational resources coupled with dynamic conditions that necessitate real-time adaptation of model execution. However, these two challenges are often addressed separately in existing research. This article presents LeapNet, which aims to address both challenges simultaneously. It comprises two versions: LeapNet-1 and LeapNet-2. LeapNet-1 employs dynamic layer routing to counteract adaptive adversarial-example attacks and reduce computational redundancy. Building upon LeapNet-1, LeapNet-2 further adapts its layer routing configurations in real time to meet the frame processing rate requirements under dynamic conditions while maintaining defense performance. Extensive experiments on various representative datasets, neural network models, and adaptive attacks demonstrate the superiority of LeapNet over existing defense methods. On-road tests with a real-time car-borne traffic sign recognition system validate its effectiveness in maintaining frame processing rate under dynamic conditions.

The rapid development of intelligent machine learning and 5G/6G has led to smart terminal devices for various applications like auto-driving, AR, and smart farms. However, these applications have strict Quality of Service (QoS) requirements, exceeding the capabilities of mobile devices due to limited resources. Mobile edge computing offers a promising solution by enabling task offloading to nearby edge servers with better computing capabilities and a stable energy supply. For mobile edge systems, minimizing overall cost while ensuring QoS for mobile users is a challenging problem to solve effectively.In this paper, we first formulate the performance model for mobile edge systems with real data from deep learning scenarios. Then, we formulate the resource management in mobile edge system problem as a non-convex fractional programming problem with multiple coupled variables to minimize the latency and energy consumption while meeting QoS requirements. To solve the non-convex problem, a novel fractional programming technique is proposed to decouple the variables by considering fractional transformation to greatly reduce the complexity. Then we achieve the jointly optimal solution of the CPU frequency and offloading strategy with successive convex approximation, and Karush-Kuhn-Tucker (KKT) conditions. The experimental results show that our proposed algorithm can achieve cost-effective solutions while meeting QoS requirements over baselines.