analiza aiqbench
ERASE targets indirect prompt injection in multimodal models, where malicious instructions hidden in images, audio, or video can hijack a model's behavior. Its main claim is that it blocks attacks without the usual cost to task accuracy, and the reported gains are concrete, such as 60.1% versus 33.1% attacked-input accuracy on 3B audio. These are single-benchmark results from a preprint with code not yet released, so readers should wait for independent replication before treating the defense as robust against adaptive attackers.
Erase the Injected Region: Multimodal Prompt-Injection Defense with On-Policy Self-Distillation
arXiv:2610.41307v1 Announce Type: cross Abstract: Multimodal indirect prompt injection embeds malicious instructions in external media that also contains information needed for the user's task. Existing defenses can block the attack yet impair the model's use of that information, leaving a gap between attack defense and task preservation. We propose ERASE, an on-policy self-distillation method that constructs its reference by erasing the injected region in place. Modality-appropriate blank content replaces the injection while the surrounding media is preserved. A student reading attacked media learns from a frozen base model reading the erased view through full-distribution token-level feedback on its own rollouts. Across image, audio, and video on two model sizes, ERASE records near-zero observed attack success on the evaluated benchmark while retaining task utility close to the undefended clean-input level. On 3B audio, ERASE reaches 60.1% attacked-input accuracy compared with ARGUS's 33.1%, within 0.5 points of the undefended clean-input level. Experiments show that ERASE reduces injection dependence from 6.26 to 0.0095 on 3B audio while staying close to the base model on the erased input. Out-of-domain accuracy also stays near the base level. Our code will be made publicly available upon acceptance.
Abstract:Multimodal indirect prompt injection embeds malicious instructions in external media that also contains information needed for the user's task. Existing defenses can block the attack yet impair the model's use of that information, leaving a gap between attack defense and task preservation. We propose ERASE, an on-policy self-distillation method that constructs its reference by erasing the injected region in place. Modality-appropriate blank content replaces the injection while the surrounding media is preserved. A student reading attacked media learns from a frozen base model reading the erased view through full-distribution token-level feedback on its own rollouts. Across image, audio, and video on two model sizes, ERASE records near-zero observed attack success on the evaluated benchmark while retaining task utility close to the undefended clean-input level. On 3B audio, ERASE reaches 60.1% attacked-input accuracy compared with ARGUS's 33.1%, within 0.5 points of the undefended clean-input level. Experiments show that ERASE reduces injection dependence from 6.26 to 0.0095 on 3B audio while staying close to the base model on the erased input. Out-of-domain accuracy also stays near the base level. Our code will be made publicly available upon acceptance.
Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
arXiv:2610.41307 [cs.CR]
(or arXiv:2610.41307v1 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2610.41307 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 284 cuvinte