Creating, Using, Misusing, and Detecting Deep Fakes

Farid, Hany

doi:10.54501/jots.v1i4.56

Cited by 35 publications

(14 citation statements)

References 51 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…If one's concern is the use of generative AI for deceit, it seems exceedingly unlikely that people who are trying to deceive will opt in to selfdisclosure. Methods for detecting synthetic media via machine learning [14], crowdsourcing [14], and digital forensics [8] offer possibilities for active classification, although as generative AI technology continues to advance, it will likely get harder and harder to detect generated content. More generally, the question of who gets to decide what content gets to count as AI-generated is central, and for labels to be effective, it is critical for the process to be percieved as legitmate [22].…”

Section: Discussion Limitations and Future Workmentioning

confidence: 99%

What label should be applied to content produced by generative AI?

Epstein

Arechar²,

Rand

2023

Preprint

View full text Add to dashboard Cite

With the rise of generative AI, there has been a recent push for disclosing whether content is produced by AI. However, it is not clear what the right terms are to use for such labels. Even the answer to the question of what counts as AI-generated is not clear, since there exists a continuum of different types of content with varying levels of algorithmic intervention. Furthermore, it is unclear how members of the general public understand various terms used in conjunction with generative AI. In this paper, we therefore investigate how the public understands the mapping between ten potential labeling terms and 15 different types of content that vary in the extent to which they are AI generated, and the extent to which they are potentially misleading. To do so, we conduct a study using N=2038 Americans (quota-sampled to the national distribution on age, gender, ethnicity, and geographic region), as well as replications with samples from Mexico and Brazil (with translated terms, N=825 and 839, respectively; quota-matched to the national distributions on age and gender). Participants are randomly assigned one of the ten terms, and then indicate which pieces of content they would consider that term as applying to. We ask which terms are most effective in terms of classification accuracy (i.e. are applied to content that should be labeled and not applied to content that should not be labeled). We find that different terms satisfy different labeling goals. Across all three countries, ``AI Generated'' is one of the terms most consistently associated by participants with content that is generated using AI, regardless of whether that content is misleading - and thus is not consistently associate with misleading content. Conversely, ``Manipulated'' and ``Not real'' were two of the terms most consistently ascribed to content that is misleading, regardless of whether the misleading content was generated using AI - and thus is not consistently associated with content that was AI generated. Interestingly, ``Artificial'' does fairly well at both classification tasks in the US and Brazil (although less so in Mexico). Finally, while labels using any of the ten terms evoked at least some amount of negative feeling towards the content and the poster among US participants, numerous terms (including ``AI generated'' and ``Artificial'') had positive associations on average for participants in Brazil and Mexico. These results have important implications for how and where generative AI disclosure is implemented, and suggest that platforms and civil society must decide carefully what their objective is for such disclosures.

show abstract

Section: Discussion Limitations and Future Workmentioning

confidence: 99%

What label should be applied to content produced by generative AI?

Epstein

Arechar²,

Rand

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…Our training used astic gradient descent with 0.9 momening rate schedule (decreasing the learny 8 epochs). Polyak averaging [13] was nal model used at inference time.…”

Section: Methodsmentioning

confidence: 99%

“…A "deep fake" of a sensor output is an artificially constructed "realistic looking" fake sensor output obtained using a deep generator architecture, such that it is not possible to distinguish it from the real sensor data [12] . Deep fakes have been successfully constructed for different types of sensor data including voice, video, and image [13] . With the possibility of having deep fakes being used as inputs to an AI inference system, the foundational assumption of the AI algorithms powering these "smart systems" have been demolished.…”

Section: Introductionmentioning

confidence: 99%

Deep Feature Learning for Intrinsic Signature Based Camera Discrimination

Banerjee

Doppalapudi

Pasiliao

et al. 2022

Big Data Min. Anal.

View full text Add to dashboard Cite

In this paper we consider the problem of "end-to-end" digital camera identification by considering sequence of images obtained from the cameras. The problem of digital camera identification is harder than the problem of identifying its analog counterpart since the process of analog to digital conversion smooths out the intrinsic noise in the analog signal. However it is known that identifying a digital camera is possible by analyzing the camera's intrinsic sensor artifacts that are introduced into the images/videos during the process of photo/video capture. It is known that such methods are computationally intensive requiring expensive pre-processing steps. In this paper we propose an end-to-end deep feature learning framework for identifying cameras using images obtained from them. We conduct experiments using three custom datasets: the first containing two cameras in an indoor environment where each camera may observe different scenes having no overlapping features, the second containing images from four cameras in an outdoor setting but where each camera observes scenes having overlapping features and the third containing images from two cameras observing the same checkerboard pattern in an indoor setting. Our results show that it is possible to capture the intrinsic hardware signature of the cameras using deep feature representations in an end-to-end framework. These deep feature maps can in turn be used to disambiguate the cameras from each another. Our system is end-to-end, requires no complicated pre-processing steps and the trained model is computationally efficient during testing, paving a way to have near instantaneous decisions for the problem of digital camera identification in production environments. Finally we present comparisons against the current state-of-the-art in digital camera identification which clearly establishes the superiority of the end-to-end solution.

show abstract

“…Continued consumer interest in deepfakes is reflected in the proliferation of dedicated deepfake sites and forums, often depicting celebrity targets. While deepfakes can be used in beneficial ways for accessibility and creativity [19,26], abuse potential has increased in recent years as the technology has advanced in sophistication and availability [12,34,53,80]. Deepfakes can be weaponized and used for malicious purposes, including financial fraud, disinformation dissemination, cyberbullying, and sexual extortion ("sextortion") [4,26].…”

Section: Introductionmentioning

confidence: 99%

Non-Consensual Synthetic Intimate Imagery: Prevalence, Attitudes, and Knowledge in 10 Countries

Umbach,

Henry,

Beard

et al. 2024

Proceedings of the CHI Conference on Human Factors in Computing Systems

View full text Add to dashboard Cite

Deepfake technologies have become ubiquitous, "democratizing" the ability to manipulate photos and videos. One popular use of deepfake technology is the creation of sexually explicit content, which can then be posted and shared widely on the internet. Drawing on a survey of over 16,000 respondents in 10 different countries, this article examines attitudes and behaviors related to "deepfake pornography" as a specific form of non-consensual synthetic intimate imagery (NSII). Our study found that deepfake pornography behaviors were considered harmful by respondents, despite nascent societal awareness. Regarding the prevalence of deepfake porn victimization and perpetration, 2.2% of all respondents indicated personal victimization, and 1.8% all of respondents indicated perpetration behaviors. Respondents from countries with specific legislation still reported perpetration and victimization experiences, suggesting NSII laws are inadequate to deter perpetration. Approaches to prevent and reduce harms may include digital literacy education, as well as enforced platform policies, practices, and tools which better detect, prevent, and respond to NSII content.

show abstract

Creating, Using, Misusing, and Detecting Deep Fakes

Cited by 35 publications

References 51 publications

What label should be applied to content produced by generative AI?

What label should be applied to content produced by generative AI?

Deep Feature Learning for Intrinsic Signature Based Camera Discrimination

Non-Consensual Synthetic Intimate Imagery: Prevalence, Attitudes, and Knowledge in 10 Countries

Contact Info

Product

Resources

About