Securing the Edge: A New Anti-Jamming Strategy for Collaborative AI Inference
As artificial intelligence moves to the network's edge, a critical vulnerability has emerged. A new research paper proposes a robust defense strategy for device-edge collaborative inference, a paradigm where deep neural networks (DNNs) are split between mobile devices and edge servers. The work directly addresses the threat of malicious jamming attacks that can cripple these systems by disrupting the transmission of intermediate feature data, which is essential for completing the AI task.
The study, detailed in the preprint "Anti-Jamming Collaborative Inference via Joint Optimization of Computation Resource Allocation, Transmit Power, and DNN Partitioning" (arXiv:2603.02579v1), formulates a novel optimization framework. It aims to maximize a composite Revenue of Delay and Accuracy (RDA) metric while ensuring inference accuracy and adhering to computing resource limits. This is achieved by simultaneously optimizing three key system parameters: computation resource allocation on the edge, the wireless transmit power of devices, and the precise DNN partitioning point.
Overcoming a Critical System Vulnerability
The core challenge in collaborative inference is the wireless link between the device and the edge server. An adversary can deploy a jammer to corrupt or block the transmission of intermediate features, leading to catastrophic drops in inference accuracy and increased latency. The researchers first conducted a rigorous analysis to model the combined impact of jamming interference and the chosen DNN split point on the final system performance. This analysis forms the foundation for their defensive optimization strategy.
To counter the jammer, the system must intelligently adapt. Allocating more edge computing resources or increasing the device's transmit power can help overcome interference, but these actions consume limited energy and infrastructure. Similarly, choosing where to partition the DNN model—executing more layers on the device or offloading them earlier—fundamentally changes the volume and vulnerability of the transmitted data. The proposed scheme dynamically balances these levers.
A Sophisticated Alternating Optimization Algorithm
The joint optimization problem is classified as a challenging mixed-integer nonlinear programming (MINLP) problem. To solve it efficiently, the authors designed an alternating optimization-based algorithm. This approach decomposes the complex problem into three more manageable subproblems:
- Resource & Power Allocation: Solved using Karush-Kuhn-Tucker (KKT) conditions and convex optimization methods for optimal efficiency.
- DNN Partitioning: Addressed with a quantum genetic algorithm (QGA), a metaheuristic well-suited for navigating the discrete, combinatorial search space of potential partition points within the neural network architecture.
By iteratively solving these subproblems, the algorithm converges on a coordinated strategy that robustly defends against jamming.
Validated Performance and Key Advantages
Extensive simulation results demonstrate the superiority of the proposed joint optimization scheme. It consistently outperforms existing baseline methods in terms of the comprehensive RDA metric, proving its effectiveness in maintaining high inference accuracy and low latency even under active jamming attacks. This represents a significant step forward in securing mission-critical edge AI applications in adversarial environments.
Why This Research Matters for Edge AI
- Security is Paramount: As collaborative inference becomes mainstream for applications like autonomous vehicles and industrial IoT, protecting the data pipeline from intentional disruption is non-negotiable. This work provides a foundational defensive framework.
- Holistic System Optimization: The approach moves beyond isolated fixes, showing that security, latency, and accuracy must be co-optimized by considering compute, communication, and model architecture simultaneously.
- Practical Algorithm Design: The hybrid algorithm combining classical optimization (KKT) with advanced metaheuristics (QGA) offers a practical blueprint for solving real-world, complex engineering problems in wireless AI systems.