Original title: Feasible but Not Safe: Constraint Violations and Report-Channel Attacks in Learned Cell-Free ISAC Association
Mehdi Zafari, Iman Mohammadi, A. Lee Swindlehurst · September 2026
Modern wireless networks try to connect users and detect radar targets at the same time using spread-out access points. Because managing these connections quickly is difficult, researchers use machine learning models called graph neural networks to make instant scheduling choices. These networks learn patterns by guessing decisions in a single step, but they treat critical operational rules merely as soft learning hints. The authors examined this using their ASSENT algorithm. They found that even when the model scores high prediction accuracy, it often breaks hard network rules, creating invalid schedules. To fix this, they used a simple greedy repair step that adjusts the model's choices to satisfy all rules with very little loss in performance. However, they also found that one dishonest access point sending fake data can easily trigger rule violations. Repair steps fix some lies, while a quick cross-checking method catches the rest.
Imagine organizing a school dance with rules like 'no more than 50 students in the gym' and 'each chaperone must get a 10-minute break'. You have a volunteer assignment app that uses past dances to predict who should work where. The app might assign the right number of people overall, but it often puts too many in the gym or forgets breaks, breaking the rules. Here, the network scheduler is like the app, and the rules are hard constraints. High accuracy doesn't mean it follows all rules, and a volunteer lying about their availability can make it worse, but some lies can be spotted by comparing what different volunteers say.
This affects wireless systems that handle both communication and sensing, like advanced networks. If schedulers break essential rules, the system could fail or be compromised, so engineers must test for feasibility, not just prediction accuracy.
The study uses the ASSENT algorithm as an example, so findings may be specific to that context. Feasibility projection and detection methods are proposed but may not cover all possible attack types or system configurations.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Feasible but Not Safe: Constraint Violations and Report-Channel Attacks in Learned Cell-Free ISAC Association — Mehdi Zafari, Iman Mohammadi, A. Lee Swindlehurst
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.