Show HN: Neural-hash-collider – Find target hash collisions for NeuralHash (github.com)

Article summary

The article presents a tool called NeuralHash Collider, which can find target hash collisions for Apple's NeuralHash perceptual hash function. This function is used to scan for child sexual abuse material (CSAM) on devices. The tool uses a standard adversarial attack against NeuralHash, which involves replacing the thresholding step with a differentiable approximation. The author shares the code to generate adversarial examples, which can potentially trigger false positives in CSAM detection.

Main themes

• NeuralHash collisions
• CSAM detection
• Adversarial attacks
• False positives
• Surveillance and censorship
• Apple's security measures
• Child protection and safety

What commenters say

• The existence of this tool undermines the effectiveness of Apple's CSAM detection system.
• The system's use of a second, private perceptual hash and human review provides additional layers of protection against false positives.
• The ability to generate adversarial images raises concerns about the potential for false accusations and account suspensions.
• Some commenters argue that the system is flawed and will inevitably be bypassed, while others believe that the additional security measures make it more robust.
• There are concerns about the potential for the system to be used as a backdoor for surveillance or censorship.
• The use of NeuralHash for CSAM detection may not be effective in catching child abusers, as it may be easily evaded or trigger false positives.
• The fact that the hashing algorithm is not final and may be changed does not address the underlying concerns about the system's design and potential vulnerabilities.