Apple child-abuse scanner has major weakness, experts warn

Apple Flaw

Scientists have discovered a vulnerability in iOS’s built-in hash function, raising worries about the system’s integrity. The vulnerability affects Apple’s NeuralHash hashing technology, which enables it to search for precise matches of known child abuse photos without really owning or seeing the images.

What is wrong with Apple Child-abuse Scanner

An iOS developer named Asuhariet Ygvar released code for a Python version of NeuralHash on GitHub on Tuesday. How to extract NeuralMatch files from a current macOS or iOS build is also documented on GitHub.

According to early testing, it can withstand picture scaling and compression but not cropping or rotation, Ygvar said on Reddit. It should help us better understand NeuralHash and its possible problems before it’s activated on all iOS devices.

They created a collision: two pictures with the same hash

After the code was made public, additional serious vulnerabilities were found. Cory Cornelius created an algorithm collision: two pictures with the same hash. If the results stand up, it will be a major flaw in Apple’s new encryption.

iOS devices now have a new mechanism for preventing child abuse images. As part of the new approach, iOS will compare local files against hashes of child abuse images compiled by the National Center for Missing and Exploited Children (NCMEC). The technology limits scans to iCloud pictures and sets a threshold of 30 matches discovered before an alert is issued. However, privacy activists are still worried about the consequences of searching local storage for illicit content, and the latest discovery has heightened worries.

While the collision is substantial, exploiting it in reality would be difficult. Collision attacks discover identical inputs that generate the same hash. In Apple’s system, this means creating an image that triggers CSAM alerts even if it isn’t a CSAM picture. Accessing the NCMEC hash database would involve creating over 30 colliding pictures and sneaking them onto the target’s phone. Only Apple and NCMEC would get an alert, which would readily identify the pictures as false positives.

The shortcoming will only increase criticism of the new reporting method. A proof-of-concept collision, like the SHA-1 collision in 2017, generally means the underlying method should be abandoned. Apple may update the algorithm or make more gradual adjustments to minimize future threats. Requests for response from Apple went unanswered.

A wider range of demands for Apple to abandon its plans for on-device scanning have grown in the weeks after the announcement. A petition titled “Tell Apple: Don’t Scan Our Phones” was started by the Electronic Frontier Foundation on Tuesday. It had almost 1,700 signatures at press time.

Get other technological news at SCIENCE AND TECHNOLOGY here!