Securing the Edge: New Framework Counters Jamming Threats in Collaborative AI Inference
As artificial intelligence moves to the network's edge, a critical vulnerability has emerged: the wireless link between devices and servers is a prime target for disruption. New research proposes a robust optimization framework to safeguard device-edge collaborative inference, a paradigm where a deep neural network (DNN) is split between a resource-constrained device and a powerful edge server. The work, detailed in the preprint "Anti-Jamming Collaborative Inference for DNN Partitioning-Based Device-Edge Co-Inference," directly addresses how malicious jamming of the intermediate data transmission catastrophically degrades inference accuracy and system latency.
The proposed system models a scenario where a malicious jammer actively interferes with the channel. The core challenge is to intelligently decide where to split the DNN, how much computational resource to allocate on the edge, and what transmit power the device should use—all under the constant threat of jamming. The goal is to maximize a novel Revenue of Delay and Accuracy (RDA) metric, which holistically balances speed and correctness against the energy and resource costs of defensive measures.
Analyzing the Threat and Engineering a Defense
The researchers first established a foundational analysis, using data regression to quantify the precise impact of jamming interference and different DNN partitioning points on the final inference performance. This analysis informed the construction of a complex optimization problem with integer and continuous variables, framed as a mixed-integer nonlinear programming (MINLP) challenge. Solving this directly is computationally prohibitive, especially for real-time adaptation at the edge.
To achieve a practical solution, the team devised an efficient alternating optimization-based algorithm. This method breaks the monolithic problem into three manageable subproblems:
- Resource Allocation: Solving for optimal computational resource distribution using Karush-Kuhn-Tucker (KKT) conditions.
- Power Control: Determining the device's optimal transmit power through convex optimization methods to combat jamming.
- Partition Point Selection: Identifying the best layer to split the DNN using a quantum genetic algorithm (QGA), which leverages quantum computing principles for efficient search in a large, discrete space.
Why This New Anti-Jamming Framework Matters
Extensive simulation results demonstrate that the joint optimization strategy significantly outperforms existing baseline methods in terms of the overall RDA metric. This research is not merely theoretical; it provides a essential blueprint for building resilient, real-world AI systems. As autonomous vehicles, industrial IoT, and smart city applications rely more on split-second, edge-based AI, guaranteeing reliable performance under adversarial conditions becomes paramount.
- Resilient Edge AI: The framework enables collaborative inference to function reliably even when the wireless channel is under active attack, moving beyond naive retransmission schemes.
- Holistic System Optimization: It pioneers the joint optimization of DNN architecture, communication parameters, and compute resources as a single anti-jamming strategy, maximizing the trade-off between accuracy, latency, and energy use.
- Practical Algorithm Design: The decomposition into solvable subproblems and the use of a QGA for partition search offer a computationally feasible path for implementation in real edge computing stacks.
This work, available on arXiv under the identifier 2603.02579v1, marks a significant step toward secure and efficient distributed AI. It ensures that the performance gains of device-edge co-inference are not undone by a simple, yet potent, wireless threat, paving the way for more robust and trustworthy edge intelligence ecosystems.