Securing the AI Edge: New Framework Counters Jamming Attacks in Collaborative Inference
As deep neural network (DNN) inference increasingly shifts to resource-constrained devices, a new security threat is emerging. A novel research framework proposes a robust solution to a critical vulnerability in device-edge collaborative inference, where malicious jamming attacks on transmitted data can cripple AI performance. The work, detailed in the paper "Anti-Jamming Collaborative Inference for DNN Partitioning-Based Device-Edge Co-Inference," introduces a joint optimization strategy to maximize system efficiency while defending against these disruptive attacks.
In the collaborative inference paradigm, a DNN model is split, or partitioned, with one segment running on a wireless device and the other on a more powerful edge server. The intermediate feature data must be transmitted between them, creating a point of failure. A malicious jammer can target this wireless link, corrupting the data and causing a significant drop in inference accuracy. This research directly addresses this threat by modeling an anti-jamming system that intelligently adapts its operation.
Optimizing Against Adversarial Disruption
The core of the proposed system is a multi-variable optimization problem. The researchers first conducted a regression analysis to quantify the precise impact of both jamming intensity and the chosen DNN partitioning point on final accuracy. Their objective is to maximize a composite Revenue of Delay and Accuracy (RDA) metric, which balances speed and correctness, under strict constraints for minimum accuracy and available computing resources.
To achieve this, the framework simultaneously optimizes three key parameters: the allocation of computing resources at the edge, the transmit power of the wireless devices, and the optimal DNN partitioning strategy. This creates a challenging mixed-integer nonlinear programming (MINLP) problem. The team developed an efficient alternating optimization-based algorithm to solve it by breaking it down into three manageable subproblems.
A Hybrid Algorithmic Solution
Each subproblem is tackled with a specialized method, showcasing a hybrid approach. Solutions for resource allocation and transmit power are derived analytically using Karush-Kuhn-Tucker (KKT) conditions and convex optimization techniques. The most complex subproblem—finding the optimal partition point in the DNN—is solved using a quantum genetic algorithm (QGA), a metaheuristic that combines quantum computing principles with evolutionary search for navigating large, discrete solution spaces efficiently.
Extensive simulation results validate the framework's superiority. The proposed joint optimization scheme demonstrably outperforms existing baseline methods in terms of the overall RDA performance, proving its effectiveness in maintaining reliable and efficient AI inference even under adversarial jamming conditions.
Why This Matters for Edge AI
- Critical Security Gap Addressed: This research identifies and mitigates a tangible security risk in the foundational device-edge co-inference model, which is essential for scaling AI to IoT and mobile applications.
- Holistic System Optimization: The work moves beyond simple jamming mitigation, integrating defense directly into a performance-maximizing framework that jointly manages compute, communication, and model architecture.
- Practical Algorithm Design: The combination of classical optimization (KKT, convex methods) with advanced metaheuristics (Quantum GA) provides a scalable and effective blueprint for solving complex, real-world AI deployment problems.