Notes: Available from: https://arxiv.org/abs/2108.11299
Abstract. A key concept towards reliable, robust, and safe AI systems is the idea to implement fallback strategies when predictions of the AI cannot be trusted. Certifiers for neural networks have made great progress towards provable robustness guarantees against evasion attacks using adversarial examples. These methods guarantee for some predictions that a certain class of manipulations or attacks could not have changed the outcome. For the remaining predictions without guarantees, the method abstains from making a prediction and a fallback strategy needs to be invoked, which is typically more costly, less accurate, or even involves a human operator. While this is a key concept towards safe and secure AI, we show for the first time that this strategy comes with its own security risks, as such fallback strategies can be deliberately triggered by an adversary. In particular, we conduct the first systematic analysis of training-time attacks against certifiers in practical application pipelines, identifying new threat vectors that can be exploited to degrade the overall system. Using these insights, we design two backdoor attacks against network certifiers, which can drastically reduce certified robustness. For example, adding 1% poisoned data during training is sufficient to reduce certified robustness by up to 95 percentage points, effectively rendering the certifier useless. We analyze how such novel attacks can compromise the overall system’s integrity or availability. Our extensive experiments across multiple datasets, model architectures, and certifiers demonstrate the wide applicability of these attacks. A first investigation into potential defenses shows that current approaches are insufficient to mitigate the issue, highlighting the need for new, more specific solutions.