Digital Media Distortion Tool for Deepfake Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Public-facing organizations are vulnerable to deepfakes due to the abundance of original media used for training generative machine learning algorithms, which can be exploited to create disinformation, and existing detection methods are inadequate and constantly evolving, making it difficult to protect against such threats.
Innovation Solution
A digital media distortion tool that intercepts original media before transmission, applies a media modification process by replacing specific data elements with new values, reducing the accuracy of generative machine learning algorithms trained on the modified media, thereby reducing the number and detectability of deepfakes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If original media is released to the public, then organizational information dissemination is improved, but vulnerability to deepfake generation increases
Solution Approach 1:
The system applies preliminary distortion to original media before public release by modifying pixel values through a distortion function. This preliminary action ensures that when the media is later used for training generative models, the distorted characteristics reduce the model's ability to create accurate deepfakes, thus protecting the organization in advance without preventing information dissemination.
2Measurement precision
If media modification is applied to reduce deepfake accuracy, then deepfake generation accuracy is reduced, but media quality is degraded
Solution Approach 1:
The distortion function selectively modifies certain pixel values while preserving others, applying different transformation intensities to different regions of the media. This local quality approach ensures that critical visual information remains intact for human perception while specific patterns are distorted enough to reduce deepfake generation accuracy.
Solution Approach 2:
The system adjusts pixel value parameters through a controlled distortion function that modifies statistical properties of the media (such as histogram distribution) without completely transforming the visual content. This parameter change approach reduces the fidelity available to generative models while maintaining acceptable quality for human consumption.
3Difficulty of detecting and measuring
If detection tools are used to identify deepfakes, then deepfake detection capability is improved, but resources and time are consumed
Solution Approach 1:
Instead of relying on post-generation detection, the system applies preliminary anti-action by distorting the original media before release. This prevents generative models from learning accurate representations in the first place, reducing the need for resource-intensive detection tools and eliminating the time lag between deepfake creation and detection.
Data Source
AI summary
An apparatus includes a processor that monitors transmissions destined for an external network, determines that a transmission includes original media associated with a subject, and intercepts the transmission before it reaches the external network. The processor generates modified media by selecting a subset of data elements of the original media and replacing a value of each data element of the subset with a new value. At least one of the subset of data elements and the set of new values is chosen such that an accuracy metric calculated for a first generative algorithm, trained to generate synthetic representations of the subject based on modified media, is less than, by a given factor, the accuracy metric calculated for a second generative algorithm, trained to generate synthetic representations of the subject based on original media. The processor replaces the transmission with a new transmission that includes the modified media.