Battery-free IoT devices and medical implants get their operating power by harvesting energy from radio signals. Because cheap hardware cannot harvest energy and transmit data on the same frequency band at the same time, every communication block is split into two phases: harvest energy first for a fraction $\alpha$ of the block, then spend the remaining $(1-\alpha)$ sending data. Choosing $\alpha$ is the central resource-allocation problem in these systems, and because the choice must track fast-varying wireless channels while also deciding which layers of a neural network run on-device versus on an edge server, recent work hands it to a deep reinforcement learning policy trained end-to-end (Joint DNN partitioning and resource allocation...). This proposal asks a question nobody has asked: what happens when that policy is backdoored?
A backdoor is a hidden conditional behavior implanted during training: the model behaves normally on ordinary inputs, but when a specific trigger appears, it does something the attacker chose (BackdoorBench: A Comprehensive Benchmark of Backdoor Learning). Backdoors are well studied for image classifiers and, increasingly, for continuous regression models. But the policy that outputs the time split $\alpha \in (0,1)$ is exactly a continuous regression head attached to a safety-critical physical controller, and no existing work — on either side of this divide — has examined whether it can be quietly hijacked. The stakes are physical rather than semantic: a triggered misallocation of a few tens of percent in $\alpha$ can mean a device that harvests energy it never gets to use, or transmits with power it never harvested.
Core Claim
The claim is threefold. First, a neural time-switching allocator can be backdoored so that, when the observed channel state falls in an attacker-chosen trigger region (an SNR band or fading condition, in the spirit of the physical triggers shown to work in semantic communication by A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction), the policy outputs a time split that induces energy starvation or deadline failure, while clean-state behavior stays within a few percent of optimal. Second, the attack should succeed precisely where the physics is most helpful to the attacker: because throughput is strictly concave in $\alpha$, with a flat top near the optimum and steep flanks near the boundary conditions (Wireless-Powered Communications: Performance Analysis and Optimization), a backdoor can hide small deviations in $\alpha$ in the steep regions where the same deviation costs far more. Third, the same concavity gives us a defense no generic backdoor detector has: for the point-to-point case the optimal split $\alpha^*$ is known in closed form, so a trained policy can be audited against solver-computed ground truth rather than against data-driven statistics.
Research Program
The throughput of a harvest-then-transmit link behaves as
$$ R(\alpha) \propto (1-\alpha)\,\log_2\!\big(1 + \mathrm{SNR}(\alpha)\big), \qquad \mathrm{SNR}(\alpha) \propto \frac{\alpha}{1-\alpha}, $$
so the linear decay of the transmission window beats the logarithmic gain in power, producing the strict concavity established in Wireless-Powered Communications: Performance Analysis and Optimization and emphasized in Utility-Based Wireless Resource Allocation for Variable Rate Transmission. This curve defines the attack surface. We will build three attack variants against a reimplementation of the DRLPR-style setting of Joint DNN partitioning and resource allocation...: a starve trigger that pushes $\alpha \to 0$ so the device transmits with no harvested power; a stall trigger that pushes $\alpha \to 1$ so the transmission window collapses and deadlines are missed; and the interesting one, a gray trigger that outputs $\alpha^*(s) - \delta$ only in a trigger region chosen to sit on a steep flank of the curve, where the induced completion-rate loss is large relative to the size of the deviation.
Poisoning the policy training set will follow the localized-poisoning recipe of Backdoor Attack and Defense for Deep Regression: that work showed naive outlier poisoning fails in regression because training smooths it away, and that poisoned samples must be buffered by dense clean samples to produce sharp, surgical output deviations while keeping overall error low. We adapt this to a reinforcement-learning experience stream, which introduces a genuinely new question — whether a poisoned trajectory fragment survives credit assignment — that neither the regression-backdoor literature nor the wireless-powered MEC literature has touched. The poisoning-paradox finding in BackdoorBench: A Comprehensive Benchmark of Backdoor Learning, that excessive poisoning makes attacks more detectable, will be mapped onto a poisoning-ratio sweep for the policy.
Two extensions round out the program. In the multi-user time-division setting of Joint DNN partitioning and resource allocation..., where scaling from $N=20$ to $N=30$ users already shrinks feasible harvesting windows, we will measure whether a single triggered node that inflates its demanded transmission window can compress every clean node's schedule below its operational threshold — a cell-level denial of service launched from one compromised device, in the multi-user allocation regime studied by Placement and Resource Allocation of Wireless-Powered Multiantenna UAV for Energy-Efficient Multiuser NOMA. On the defense side, we will build a solver-anchored audit: run the convex subproblem solver (fmincon, exactly as DRLPR itself uses once $\alpha$ is fixed) over a grid of held-out channel states to produce reference $\alpha^*$ values, then test whether the trained policy's deviations from that reference are unimodal (clean) or bimodal with a spike localized to the trigger region (backdoored). This is the point-to-point anchor; for the mixed-integer multi-user case the anchor is approximate, and the open empirical question is how much evasion room that approximation buys a gray attack.
Proposed Experiments
All experiments run in the simulated wireless-powered MEC environment of Joint DNN partitioning and resource allocation...: an AlexNet workload profile on $1280 \times 720$ images (2.7 MB per input), Rayleigh-fading channels, device populations of $N=20$ and $N=30$, and delay deadlines on the inference tasks. The clean DRLPR reference behavior is anchored to that paper's reported figures: a normalized task completion rate of 1.0 within its optimal operating window, roughly 10% above the EDGEONLY heuristic and far above LOCALONLY and RANDOM, with the deep-RL training approaching stable convergence after approximately 1,050 time frames. The core metrics are clean-state normalized completion rate (stealth), attack success rate defined as the fraction of triggered frames in which a device misses its deadline or suffers energy outage (efficacy), the deviation $|\pi(s) - \alpha^*(s)|$ from the solver reference (forensics), and detection ROC-AUC for each defense. The decisive check is a stealth-evasion frontier: sweep the gray-attack deviation $\delta$ and the trigger-region rarity, and plot the completion-rate loss inflicted against the detection AUC of every defense. Concavity predicts a favorable region for the attacker where loss is first-order in $\delta$ on steep flanks but the deviation signal available to a data-driven detector is not; whether that region is nonempty in practice decides the proposal's significance. A second decisive check is the uniform-versus-localized poisoning ablation, which should replicate the smoothing failure predicted by Backdoor Attack and Defense for Deep Regression in the policy setting, confirming that our attack's localized buffering is the operative ingredient rather than generic label noise.
Baselines
The do-nothing reference is the clean DRLPR policy itself, with its reported 1.0 normalized completion rate inside the optimal window and roughly 10% margin over EDGEONLY (Joint DNN partitioning and resource allocation...); a successful attack must hold clean-state completion within a tolerance comparable to the 1–2% utility drop tolerated by the split-learning backdoor of Chronic Poisoning: Backdoor Attack against Split Learning while causing large triggered loss, and any defense must not degrade clean completion beyond that same band. The strongest attack competitor is the localized regression poisoning of Backdoor Attack and Defense for Deep Regression, which reported attack success above 95% at 5–8% poisoning on its own tasks; we will transplant it wholesale to the policy's training stream — same poisoning budget, same trigger definition, same tuning effort — and our contribution counts only if the channel-state trigger and the steep-flank targeting add measurable stealth or efficacy beyond it. We will also run overt RF interference as a control: elevate the noise floor on the same cell and show that a simple SINR monitor flags it immediately while the backdoor remains statistically invisible on clean traffic, matching the near-zero clean-state footprint reported for channel-triggered backdoors (clean PSNR differences below 0.5% at a 1% poisoning ratio in A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction). On the defense side, the strongest existing competitor is DRMGuard from Defending Deep Regression Models against Backdoor Attacks, which reported an average identification accuracy of 95% on MPIIFaceGaze and ROC-AUC of 1.000 where classification defenses scored about 0.5; we will run it on the policy's activations with its stated budget of roughly 10% of the training data and comparable per-model compute (its reported overhead is about 16 minutes per model), and our solver-anchored audit beats it only if it detects gray and rare-trigger attacks that DRMGuard's feature-angle statistics (concentrated angles with RAV below 0.1 in that paper's backdoored models) miss. The key ablation is the audit stripped of its solver anchor — an output-space anomaly detector with identical architecture and budget — which isolates whether the physics-derived reference $\alpha^*$, rather than generic anomaly detection, is what buys the detection power.
Why not just…?
Why not just jam the network? Overt interference is the obvious way to hurt a wireless-powered cell, but it fails on both stealth and sometimes even efficacy. Jamming degrades the signal-to-interference ratio for every user at once, which standard deep-RL networking stacks are explicitly built to detect and route around (QoS and Jamming-Aware Wireless Networking Using Deep Reinforcement Learning), and ambient RF interference can even be harvested as a free energy source in interference-based energy-harvesting relays (On the Performance of Interference-Based Energy-Harvesting-Enabled Wireless AF Relaying Communication Systems). The backdoor, by contrast, keeps clean behavior indistinguishable — the closest evidence being clean PSNR differences below 0.5% in A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction and a 1–2% utility drop in Chronic Poisoning — and strikes only the triggered regime. The jamming control in our experiments settles this empirically.
Why not just drop the neural policy and use a convex solver? This is the strongest objection, and it deserves an honest answer. In a static point-to-point link the optimal $\alpha^*$ follows from the concavity analysis of Wireless-Powered Communications: Performance Analysis and Optimization, and no learned policy is needed. But the regime where DRL is actually being deployed is the joint one: discrete partition points, per-device power and frequency, time-varying channels, and deadline pressure, a mixed-integer nonlinear program in which Joint DNN partitioning and resource allocation... showed static heuristics lose roughly 10% (EDGEONLY) or fail outright (LOCALONLY, RANDOM), and which iterative solvers cannot re-solve in real time. The field is moving to learned allocators regardless of this proposal, so characterizing their attack surface is warranted on its face. Moreover, the honest version of this objection — "just compare the policy's output to the solver's answer at runtime" — is exactly our audit defense; the research question it cannot settle alone is whether it can be evaded by gray deviations inside solver tolerance and by triggers placed where the multi-user reference solution is itself approximate, which is precisely what the stealth-evasion frontier measures.
Why not just run existing backdoor defenses? Classification-era defenses fail here by construction: Defending Deep Regression Models against Backdoor Attacks showed Neural Cleanse-style detection collapses to ROC-AUC around 0.5 on continuous regression outputs, and pruning-based defenses inherit the fragility documented across 8,000 attack-defense pairs in BackdoorBench, where defenses that hold on one backbone fail on another. The real threat is DRMGuard, which is why it is our primary defense baseline rather than a strawman. Our claim against it is narrow and testable: DRMGuard's statistics are learned from data, so a trigger region that is rare in the audit set and deviations tuned to the steep flanks of the concave surface should stress it, whereas the solver anchor gives ground truth exactly where the physics is hardest. If DRMGuard matches the solver-anchored audit on every point of the frontier, the defense contribution of this proposal fails — and the experiments are designed so that we would know.
Decisive Checks
The proposal succeeds if the following hold in the simulator: the gray attack inflicts a completion-rate loss several times larger than its clean-state footprint while evading output-space anomaly detection; the loss-to-deviation ratio is measurably higher on steep flanks than near the flat optimum, confirming the concavity mechanism; one triggered node measurably degrades clean nodes' completion in the $N=30$ time-division cell; and the solver-anchored audit detects the gray attack with ROC-AUC materially above DRMGuard's on the mixed-integer multi-user setting where the anchor is approximate. It fails, informatively, if runtime solver comparison alone closes the problem — in which case the contribution collapses to a well-documented negative result with a cheap, deployable mitigation.
References
- Joint DNN partitioning and resource allocation for completion rate maximization of delay-aware DNN inference tasks in wireless powered mobile edge computing
- BackdoorBench: A Comprehensive Benchmark of Backdoor Learning
- A Channel-Triggered Backdoor Attack on Wireless Semantic Image Reconstruction
- Wireless Powered Communications: Performance Analysis and Optimization
- Utility-Based Wireless Resource Allocation for Variable Rate Transmission
- Backdoor Attack and Defense for Deep Regression
- Placement and Resource Allocation of Wireless-Powered Multiantenna UAV for Energy-Efficient Multiuser NOMA
- Chronic Poisoning: Backdoor Attack against Split Learning
- DRMGuard: Defending Deep Regression Models against Backdoor Attacks
- QoS and Jamming-Aware Wireless Networking Using Deep Reinforcement Learning
- On the Performance of Interference-Based Energy-Harvesting-Enabled Wireless AF Relaying Communication Systems