Detecting Deepfakes Without Seeing Any
Tal Reiss, Bar Cavia, Yedid Hoshen
TL;DR
This paper proposes a zero-shot deepfake detection method called FACTOR that uses fact checking to verify claimed facts in media, effectively identifying unseen attack types without training on deepfakes.
Contribution
Introduces a training-free fact checking approach for deepfake detection that generalizes to zero-day attacks by verifying claimed facts against observed media.
Findings
Outperforms state-of-the-art in zero-day deepfake detection
Effective in face swapping and audio-visual synthesis attacks
Training-free and relies on off-the-shelf features
Abstract
Deepfake attacks, malicious manipulation of media containing people, are a serious concern for society. Conventional deepfake detection methods train supervised classifiers to distinguish real media from previously encountered deepfakes. Such techniques can only detect deepfakes similar to those previously seen, but not zero-day (previously unseen) attack types. As current deepfake generation techniques are changing at a breathtaking pace, new attack types are proposed frequently, making this a major issue. Our main observations are that: i) in many effective deepfake attacks, the fake media must be accompanied by false facts i.e. claims about the identity, speech, motion, or appearance of the person. For instance, when impersonating Obama, the attacker explicitly or implicitly claims that the fake media show Obama; ii) current generative techniques cannot perfectly synthesize the false…
Peer Reviews
Decision·ICLR 2024 Conference Withdrawn Submission
1. Detecting Deepfakes is a hot topic. 2. The proposed method has a very simple formulation.
**1. Falsely claimed robustness against zero-day attacks.** The arguments 'the proposed framework is robust against zero-day attacks because it uses pre-trained feature extractor without tuning on deepfake samples' is not justified. It is essentially claiming the following impossibility for generative models: Given a feature extractor, one cannot train a generative model to make the embedding of the generated samples close to a target (real) sample. This is a very strong and likely wrong claim
1. The authors have tackled a new form of deepfake detection. While I do have some concerns about the practical nature of that setting (see weaknesses), it might still be worthwhile to think about some different ways of thinking about - what makes a data point fake. 2. The authors have made an effective use of some of the large pretrained models to solve the detection task. Their main point about potential usage of these feature spaces without needing to finetune them on any supervised dataset
1. My major concern is with the new form of problem statement for detecting deepfakes for two out of the three settings presented in the paper: (i) detecting face-swapped images (Section 5), and (ii) detecting images generated by text-to-image diffusion models. For (i), when we see DeepFakes in the wild, the attacker does not explicitly give us a well structured label that we can use as identity. Even if we somehow do, it is not clear how we will use the label to collect the reference set. My po
The methodology presented in this paper is ” simple but effective”. The authors approach the problem from the perspective of fact-checking, offering a training-free method, which is particularly intriguing. The paper is well-structured, providing clear explanations of the methods employed and the corresponding results. The authors offer detailed accounts of their experiments, substantiating the effectiveness of their approach with various types of data. The introduction of the novel deepfake d
While the authors have conducted a substantial number of experiments, there are instances where certain experiments appear to oversimplify the problem. For instance, in the deepfake detection from text-to-image, the use of diffusion models to generate data may not fully capture the complexities present in real-world scenarios of deep deception. The authors have separately validated the deep deception in three distinct aspects. However, there are occasions where these aspects may not be entirely
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDigital Media Forensic Detection · Generative Adversarial Networks and Image Synthesis · Misinformation and Its Impacts