Obfuscation Network Training via Frequency Filtering for Privacy-Preserving AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data anonymization techniques distort original data, making it unusable for machine learning algorithms and fail to conceal unexpected identification information, while conventional security cameras anonymize videos by blurring every pixel, altering critical information and leaving personal identification information intact.
Innovation Solution
A method for generating obfuscated data by filtering frequency information through high-pass and low-pass filters or band reject filters, followed by concatenation, to create obfuscated data that is recognizable by machines but not humans, while maintaining the same results as the original data for learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional anonymization techniques mask or blur identification information, then privacy protection is improved, but machine learning algorithms cannot utilize the data due to distortion
Solution Approach 1:
The patent segments data into different frequency components using Fourier transformation, separating identification information (high-frequency components) from task-relevant information (low-frequency components). This allows selective filtering of privacy-carrying frequencies while preserving learning-useful frequencies.
Solution Approach 2:
The patent changes the frequency domain parameters of the data by applying Fourier transformation, filtering specific frequency bands, and then inverse transformation. This parameter transformation in the frequency domain enables privacy protection while maintaining machine learning usability, as opposed to spatial domain masking that destroys data structure.
2Reliability
If security cameras blur every pixel changed between frames, then moving target anonymization is improved, but critical information such as facial expression becomes different from original video
Solution Approach 1:
The patent segments video data into frequency components, identifying that identification information primarily resides in high-frequency bands while task-relevant information exists in low-frequency bands. This segmentation enables selective filtering without destroying critical information.
Solution Approach 2:
The patent applies frequency domain parameter changes by filtering specific frequency bands rather than uniformly blurring spatial pixels. This selective frequency filtering preserves critical low-frequency information (facial expressions, movements) while removing high-frequency identification information.
3Reliability
If conventional techniques detect and mask identification information, then privacy protection is improved, but unexpected identification information cannot be concealed
Solution Approach 1:
The patent transforms data to the frequency domain where identification information manifests as specific frequency patterns regardless of its original form. This frequency-based approach universally captures all identification information including unexpected types, unlike detection-based methods that require prior knowledge of what to look for.
Solution Approach 2:
The patent replaces the mechanical detection-and-mask system with a frequency-based filtering system. Instead of detecting identification information in the spatial domain and applying masks, the system uses Fourier transformation to automatically separate identification frequencies from task-relevant frequencies, providing universal coverage without requiring detection algorithms.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A learning method for training an obfuscation network capable of obfuscating original data for privacy, including steps of: (a) inputting training data into the obfuscation network to filter frequency information of the training data and thus generate obfuscated data; and (b) (i) inputting the obfuscated data into a learning network to generate characteristic information by performing learning operation on the obfuscated data, (ii) generating at least one task loss by referring to (ii-1) the characteristic information and its corresponding ground truth or (ii-2) a task-specific output, generated by using the characteristic information, and its corresponding ground truth, and (iii) training at least one of the learning network and the obfuscation network through a backpropagation of the task loss.