Abstract
Recent advancements in deploying deep neural networks (DNNs) on resource-constrained devices have generated interest in input-adaptive dynamic neural networks (DyNNs). DyNNs offer more efficient inferences by using early-exit mechanisms. However, we have discovered a new vulnerability in DyNNs that could potentially compromise their efficiency. We propose EfficFrog, an adversarial attack that injects universal efficiency backdoors. By poisoning a minimal percentage of training data, EfficFrog forces backdoored DyNNs to abandon their efficiency features during inference when a trigger is present, drastically increasing computational costs and draining device resources without affecting accuracy on normal inputs.
Introduction: The Vulnerability of Dynamic Routing
Dynamic Neural Networks (DyNNs) are designed for efficiency on edge devices like smartphones or autonomous vehicles. They utilize "early-exit" mechanisms where computation stops as soon as an intermediate decision unit is sufficiently confident in a prediction. This saves battery and processing time by skipping unnecessary layers for simpler inputs.
While existing research focuses on defending against attacks that compromise model integrity (misclassification), EfficFrog targets model availability and efficiency. By forcing a DyNN to always run at maximum capacity using a subtle trigger, an adversary can launch a "Denial-of-Service" (DoS) attack on hardware, causing system lag and rapid battery depletion. This is particularly difficult to detect because there is no ground truth for the "correct" amount of computation for a given input.
Methodology: The EfficFrog Attack
EfficFrog is a two-stage adversarial attack designed to maximize computational waste:
- Adversarial Trigger Generation: The attacker optimizes a universal trigger (a small patch) to make the model's intermediate decision units as uncertain as possible. Instead of mislabeling, it pushes the model toward a uniform distribution (maximum entropy), preventing early exits.
- Backdoor Implantation: The attacker poisons a small fraction (5-15%) of the training data. The model is trained with a multi-part loss function. A Clean Loss ensures accuracy and efficiency on normal data, while an Adversarial Loss teaches the model to produce "unconfident" scores whenever the trigger is present, forcing computation to the final layer.
Comparative Landscape
EfficFrog represents a shift in the adversarial landscape. Unlike traditional backdoors like BadNets or TrojanNN, which aim for high confidence in a target class (and thus inadvertently maintain efficiency), EfficFrog specifically targets the efficiency-accuracy trade-off.
Experimental Results
The researchers evaluated EfficFrog using CIFAR-10 and Tiny ImageNet across three architectures: VGG16, MobileNet, and ResNet56. Key findings include:
- Massive Inefficiency: On triggered inputs, EfficFrog increased the number of computational blocks used by 3x to 4x. For VGG19, average block consumption jumped from ~1.0 to over 4.0.
- Stealthiness: On clean data, the backdoored model's accuracy and latency are virtually indistinguishable from a clean model, making the backdoor difficult to detect through standard performance monitoring.
- Low Poisoning Requirement: The attack remains highly effective even with only 5% of the training data poisoned.
- Efficiency-Exit Curve (EEC) Drop: The EEC score, where 1.0 represents perfect efficiency, dropped to nearly 0.5 for triggered samples, indicating a massive waste of resources.
Conclusion and Future Work
EfficFrog demonstrates that the confidence-based thresholds of DyNNs are an "Achilles' Heel" for efficiency. While the attack is highly effective and stealthy, future work must address trigger visibility and develop "efficiency-aware" defenses. Testing on real-world hardware to measure actual thermal impact and battery drain will further quantify the physical risks of these efficiency-based Denial-of-Service attacks.
Summary updated