State-of-the-art (SOTA) weight-shared SuperNets dynamically activate subnetworks at runtime, enabling robust adaptive inference under varying deployment conditions. However, we find that adversaries can take advantage of the unique training and inference paradigms of SuperNets to selectively implant backdoors that activate only within specific subnetworks, remaining dormant across billions of other subnetworks. We present VillainNet (VNET), a novel poisoning methodology that restricts backdoor activation to attacker-chosen subnetworks, tailored either to specific operational scenarios (e.g., specific vehicle speeds or weather conditions) or to specific subnetwork configurations. VNET's core innovation is a novel, distance-aware optimization process that leverages architectural and computational similarity metrics between subnetworks to ensure that backdoor activation does not occur across non-target subnetworks. This forces defenders to confront a dramatically expanded search space for backdoor detection. We show that across two SOTA SuperNets, trained on the CIFAR10 and GTSRB datasets, VNET can achieve attack success rates comparable to traditional poisoning approaches (approximately 99%), while significantly lowering the chances of attack detection, thereby stealthily hiding the attack. Consequently, defenders face increased computational burdens, requiring on average 66 (and up to 250 for highly targeted attacks) sampled subnetworks to detect the attack, implying a roughly 66-fold increase in compute cost required to test the SuperNet for backdoors.

The Android accessibility (a11y) service has been widely utilized by malware to abuse benign services.To prevent such abuse, developers need to secure a11y content access in both their apps and mobile websites.However, a misalignment of a11y protection mechanisms exists between them.Prior research has focused on attacking and defending a11y information embedded in native Android apps.However, our research found that a11y malware can retrieve app-protected a11y information in its mobile browser-rendered website counterpart, leaving mobile browser users more vulnerable to a11y attacks than app users.To help benign service developers vet this attack surface, we developed SOMBRA, an automated analysis pipeline to vet browser-side leakage of a11y information that is a11y-protected in apps.Using SOMBRA, we analyzed 294 benign services and found 29 of them deploy app-side a11y protection mechanisms to secure 256 views.SOMBRA discovered that 241, 402, 244, and 251 elements corresponding to their protected app-side views are a11y-exposed in their websites rendered by Chrome, Firefox, Brave, and Edge browsers, respectively.The leaked elements contain sensitive personal identifiable information.Finally, SOMBRA discovered that most developers do not adopt browser-side a11y protections because existing mechanisms either have ineffective protection or hinder the usability of their content.

Search engines and advertisers continuously suffer substantial financial losses from click fraud, which poses challenges to existing detection algorithms. Even more concerning, despite ongoing advancements, our understanding of click fraud remains limited, leaving room for sophisticated fraudulent techniques to bypass existing detection measures. In this study, we pivot from examining individual search requests to analyzing search sessions, defined as sequences of consecutive search queries made by the same user. We found that benign users exhibit coherent behavior patterns within these sessions, which contrast clearly with those of fraudulent actors. Specifically, legitimate users tend to conduct searches focused on a single topic at a time. In contrast, fraudsters or automated bots often exhibit diverse, illogical, and incoherent search behaviors within a session. To address this behavioral distinction, we propose CoSeC, a system designed to quantify the “incoherence index” of search sessions. CoSeC integrates literal semantic, temporal, and ad-click behavioral features to evaluate sessions' coherence quantitatively. Our evaluation of CoSeC demonstrates high efficacy, achieving a precision of 95.79% and a recall of 92.40% in identifying incoherent sessions, highlighting CoSeC's substantial potential to enhance real-world click fraud detection.