Adversarial Face Detection Protection for Deepfake-Resistant Images

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deepfake technologies, such as DeepFaceLab and FaceSwap, are highly specialized and prone to overfitting, making them immune to traditional adversarial attacks and data poisoning, thus lacking effective protection against unauthorized facial impersonation.

Innovation Solution

Implementing a computer-implemented method that embeds a barely perceptible watermark into image or video content using adversarial attacks through gradient descent to disrupt face detection mechanisms employed by deepfake creation tools, specifically using Projected Gradient Descent (PGD) with the AdaBelief optimizer to generate adversarial images that fool face detectors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional adversarial attacks or data poisoning are applied to deepfake models, then the model should be vulnerable to protection, but deepfake models are immune due to their specialized architecture and overfitting behavior

Engineering Contradiction:
Improveprotection effectivenessVSAvoidmodel specialization
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent applies preliminary action by embedding adversarial watermarks into the source content before it is used for training or inference. This proactive approach modifies the input data in advance to prevent deepfake generation, rather than attempting to defend against attacks after the model is trained. The watermarking process occurs prior to the deepfake creation pipeline, disrupting face detection mechanisms before they can function effectively.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If the model is fine-tuned per identity pair to improve accuracy, then the model becomes more specialized and effective, but this same specialization makes it immune to traditional adversarial attacks

Engineering Contradiction:
Improveidentity recognition accuracyVSAvoidattack susceptibility
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent converts the harmful effect of model specialization (which creates immunity to attacks) into a beneficial defense mechanism. By leveraging the model's tendency to overfit and its reliance on specific identity characteristics, the adversarial watermark becomes particularly effective - the model's specialization causes it to latch onto the watermark patterns, making the protection more robust rather than less effective.

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

3Reliability

If adversarial watermarks are embedded in source content, then deepfake generation is prevented, but the content may suffer from visual degradation

Engineering Contradiction:
Improvedeepfake protectionVSAvoidimage quality
Core Design Contradiction:
ReliabilityVSManufacturing precision

Solution Approach 1:

The patent applies parameter changes by carefully adjusting the adversarial watermark strength and optimization parameters to find the optimal balance between protection effectiveness and visual quality. The gradient descent optimization process modifies image parameters within constrained bounds, making subtle adjustments that disrupt face detection while minimizing perceptible degradation to the overall image quality.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4651100A1Computer-implemented method and deepfake protection system for protecting image or video content against the generation of deepfakes
Publication Date: 2025.11.19 FUNDACION CENT TECNOLOXICO DE TELECOMM DE GALICIA
  • EP4651100A1 patent drawingFigure 1A
  • EP4651100A1 patent drawingFigure 1B
  • EP4651100A1 patent drawingFigure 2

AI summary

Computer-implemented method and deepfake protection system for protecting image or video content against the generation of deepfakes. The method (100) comprises receiving (102) at least one original image (10) and iteratively: • detecting (106) faces with a first face detector (16), • calculating (108) a loss (18) using a loss function associated with the first face detector (16), • calculating (110) a gradient of the loss (18), • generating (112) an adversarial image by applying projected gradient descent (20) on the loss function using a projection operator, wherein the gradient step size is obtained using an optimizer (22) and the projection operator is a projection on a ε-ball of radius ε around the original image (10), • updating (114) a copy (14) of the original image with the adversarial image. The method obtains (116) a protected image (12) associated to each original image (10), the protected image (12) being the adversarial image obtained in the last iteration of the corresponding original image (10).