1 June 2020 in IEEE Journal of Selected Topics in Signal Processing
Deepfakes, or artificially generated audiovisual renderings, can be used to defame a public figure or influence public opinion. With the recent discovery of generative adversarial networks, an attacker using a normal desktop computer fitted with an off-the-shelf graphics processing unit can make renditions realistic enough to easily fool a human observer. Detecting deepfakes is thus becoming important for reporters, social media platforms, and the general public. In this work, we introduce simple, yet surprisingly efficient digital forensic methods for audio spoof and visual deepfake detection. Our methods combine convolutional latent representations with bidirectional recurrent structures and entropy-based cost functions. The latent representations for both audio and video are carefully chosen to extract semantically rich information from the recordings. By feeding these into a recurrent framework, we can detect both spatial and temporal signatures of deepfake renditions. The entropy-based cost functions work well in isolation as well as in context with traditional cost functions. We demonstrate our methods on the FaceForensics++ and Celeb-DF video datasets and the ASVSpoof 2019 Logical Access audio datasets, achieving new benchmarks in all categories. We also perform extensive studies to demonstrate generalization to new domains and gain further insight into the effectiveness of the new architectures.
11 November 2004 in ACM Transactions on Information and System Security
There have been a number of protocols proposed for anonymous network communication. In this paper, we investigate attacks by corrupt group members that degrade the anonymity of each protocol over time. We prove that when a particular initiator continues communication with a particular responder across path reformations, existing protocols are subject to the attack. We use this result to place an upper bound on how long existing protocols, including Crowds, Onion Routing, Hordes, Web Mixes, and DC-Net, can maintain anonymity in the face of the attacks described. This provides a basis for comparing these protocols against each other. Our results show that fully connected DC-Net is the most resilient to these attacks, but it suffers from scalability issues that keep anonymity group sizes small. We also show through simulation that the underlying topography of the DC-Net affects the resilience of the protocol: as the number of neighbors a node has increases the strength of the protocol increases, at the cost of higher communication overhead.
17 February 2011 in IEEE Transactions on Mobile Computing
While many protocols for sensor network security provide confidentiality for the content of messages, contextual information usually remains exposed. Such contextual information can be exploited by an adversary to derive sensitive information such as the locations of monitored objects and data sinks in the field. Attacks on these components can significantly undermine any network application. Existing techniques defend the leakage of location information from a limited adversary who can only observe network traffic in a small region. However, a stronger adversary, the global eavesdropper, is realistic and can defeat these existing techniques. This paper first formalizes the location privacy issues in sensor networks under this strong adversary model and computes a lower bound on the communication overhead needed for achieving a given level of location privacy. The paper then proposes two techniques to provide location privacy to monitored objects (source-location privacy)-periodic collection and source simulation-and two techniques to provide location privacy to data sinks (sink-location privacy)-sink simulation and backbone flooding. These techniques provide trade-offs between privacy, communication cost, and latency. Through analysis and simulation, we demonstrate that the proposed techniques are efficient and effective for source and sink-location privacy in sensor networks.
20 November 2020 in IEEE Transactions on Information Forensics and Security
Website Fingerprinting (WF) is a type of traffic analysis attack that enables a local passive eavesdropper to infer the victim's activity, even when the traffic is protected by a VPN or an anonymity system like Tor. Leveraging a deep-learning classifier, a WF attacker can gain over 98% accuracy on Tor traffic. In this paper, we explore a novel defense, Mockingbird, based on the idea of adversarial examples that have been shown to undermine machine-learning classifiers in other domains. Since the attacker gets to design and train his attack classifier based on the defense, we first demonstrate that at a straightforward technique for generating adversarial-example based traces fails to protect against an attacker using adversarial training for robust classification. We then propose Mockingbird, a technique for generating traces that resists adversarial training by moving randomly in the space of viable traces and not following more predictable gradients. The technique drops the accuracy of the state-of-the-art attack hardened with adversarial training from 98% to 42-58% while incurring only 58% bandwidth overhead. The attack accuracy is generally lower than state-of-the-art defenses, and much lower when considering Top-2 accuracy, while incurring lower bandwidth overheads.