Diffusion-based purification (DBP) has become a cornerstone defense against adversarial examples (AEs), regarded as robust due to its use of diffusion models (DMs) that project AEs onto the natural data manifold. We refute this core claim, theoretically proving that gradient-based attacks effectively target the DM rather than the classifier, causing DBP's outputs to align with adversarial distributions. This prompts a reassessment of DBP's robustness, accrediting it two critical factors: inaccurate gradients and improper evaluation protocols that test only a single random purification of the AE. We show that when accounting for stochasticity and resubmission risk, DBP collapses. To support this, we introduce DiffBreak, the first reliable toolkit for differentiation through DBP, eliminating gradient mismatches that previously further inflated robustness estimates. We also analyze the current defense scheme used for DBP where classification relies on a single purification, pinpointing its inherent invalidity. We provide a statistically grounded majority-vote (MV) alternative that aggregates predictions across multiple purified copies, showing partial but meaningful robustness gain. We then propose a novel adaptation of an optimization method against deepfake watermarking, crafting systemic perturbations that defeat DBP even under MV, challenging DBP's viability.

Reports regarding the misuse of Generative AI (GenAI) to create deepfakes are frequent. Defensive watermarking enables GenAI providers to hide fingerprints in their images and use them later for deepfake detection. Yet, its potential has not been fully explored. We present UnMarker— the first practical universal attack on defensive watermarking. Unlike existing attacks, UnMarker requires no detector feedback, no unrealistic knowledge of the watermarking scheme or similar models, and no advanced denoising pipelines that may not be available. Instead, being the product of an in-depth analysis of the watermarking paradigm revealing that robust schemes must construct their watermarks in the spectral amplitudes, UnMarker employs two novel adversarial optimizations to disrupt the spectra of watermarked images, erasing the watermarks. Evaluations against SOTA schemes prove UnMarker's effectiveness. It not only defeats traditional schemes while retaining superior quality compared to existing attacks but also breaks semantic watermarks that alter an image's structure, reducing the best detection rate to 43% and rendering them useless. To our knowledge, UnMarker is the first practical attack on semantic watermarks, which have been deemed the future of defensive watermarking. Our findings show that defensive watermarking is not a viable defense against deepfakes, and we urge the community to explore alternatives.

Voice authentication (VA) has recently become an integral part in numerous security-critical operations, such as bank transactions and call center conversations. The vulnerability of automatic speaker verification systems (ASVs) to spoofing attacks instigated the development of countermeasures (CMs), whose task is to differentiate between bonafide and spoofed speech. Together, ASVs and CMs form today’s VA systems and are being advertised as an impregnable access control mechanism. We develop the first practical attack on spoofing countermeasures, and demonstrate how a malicious actor may efficiently craft audio samples against these defenses. Previous adversarial attacks against VA have been mainly designed for the whitebox scenario, which assumes knowledge of the system’s internals, or requires large query and time budgets to launch target-specific attacks. When attacking a security-critical system, these assumptions do not hold. Our attack, on the other hand, targets common points of failure that all spoofing countermeasures share, making it real-time, model-agnostic, and completely blackbox without the need to interact with the target to craft the attack samples. The key message from our work is that CMs mistakenly learn to distinguish between spoofed and bonafide audio based on cues that are easily identifiable and forgeable. The effects of our attack are subtle enough to guarantee that these adversarial samples can still bypass the ASV as well and preserve their original textual contents. These properties combined make for a powerful attack that can bypass security-critical VA in its strictest form, yielding success rates of up to 99% with only 6 attempts. Finally, we perform the first targeted, over-telephony-network attack on CMs, bypassing several known challenges and enabling a variety of potential threats, given the increased use of voice biometrics in call centers. Our results call into question the security of modern VA systems and urge users to rethink their trust in them, in light of the real threat of attackers bypassing these measures to gain access to their most valuable resources.

This research is done in the context of the SliceNet project [4] that aims to extend 5G infrastructure with cognitive management of cross-domain, cross-layer network slices [1], with emphasis on Quality of Experience (QoE) for vertical industries. The provisioning of network slices with proper QoE guarantees is seen as one of the key enablers of future 5G-enabled networks. The challenge is to assess the QoE experienced by the vertical application and its users without requiring the applications or the users to measure and report QoE related metrics back to the provider. To address this challenge, we propose a method for deriving application-level QoE from network-level Quality of Service (QoS) measurements, easily accessible by the provider. In particular, we describe a PoC where QoE, perceived by application users, is estimated from low level network monitoring data, by applying cognitive methods. Our main goal is enabling the cloud provider to support the desired E2E QoE-based Service Level Agreements (SLAs), e.g. by monitoring QoS metrics within the provider's domain to optimize resource allocation through provider's actuators. Additional benefit can be achieved by applying the same technique to troubleshoot issues in the provider's infrastructure. In this work, we employed classical statistical methods to assess the relationship between the application-level QoE and the network-level QoS.