Convert Apple NeuralHash model for CSAM Detection to ONNX (github.com)

Article summary

The article describes a project that converts Apple's NeuralHash model for CSAM detection to the ONNX format, allowing for more flexibility and potential use in different frameworks. The NeuralHash model is a perceptual hashing method based on neural networks that can tolerate image resize and compression. The project provides a demo script for testing the model and discusses the potential applications and limitations of the technology. The conversion process involves several steps, including decoding the model structure and shapes, and converting the model to ONNX format.

Main themes

• NeuralHash algorithm
• ONNX format
• CSAM detection
• adversarial collisions
• perceptual hashing
• machine learning models
• human review process

What commenters say

• The NeuralHash algorithm is insufficient to defeat basic evasion techniques, such as generating invisible noise that changes the hash of an input image.
• The algorithm's vulnerability to adversarial collisions can be exploited to create images that trigger false positives or circumvent detection.
• The use of perceptual hashing systems like NeuralHash can lead to false positives and incorrect accusations, especially when combined with human review processes.
• The potential for abuse of the NeuralHash algorithm, such as creating adversarial collisions to frame individuals, is a significant concern.
• The conversion of the NeuralHash model to ONNX format can facilitate the development of more effective countermeasures and evasion techniques.
• The ONNX format is a widely-used standard for representing machine learning models, but its performance and support can vary depending on the specific use case and framework.
• The human review process involved in CSAM detection can be flawed and may lead to incorrect accusations or false positives, even with the use of advanced algorithms like NeuralHash.