Toward Sub-1 kB Identity-Preserving Face Compression: A Benchmark of Codecs, a Custom Learned Codec, and Studies of Resolution, Demographic Fairness, Recompression, and Adversarial Robustness
This paper benchmarks ten general and face-specific codecs and introduces a custom learned codec to demonstrate that while modern solutions effectively preserve facial identity under a 1024-byte budget, codec performance rankings shift drastically at 512 bytes, necessitating budget-specific deployment strategies to ensure low false-non-match rates and demographic fairness.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine holding a passport or a boarding pass. Inside that thin card, or printed as a tiny barcode, is a digital photograph of your face. This image is not just a picture for humans to see; it is a key that machines use to verify your identity. But there is a strict limit on how much space this digital key can take. In many secure systems, the entire face image must fit into a tiny pocket of memory, often no larger than a single kilobyte—about the size of a small text message. This is a severe constraint. To fit a face into such a small space, the computer must throw away almost all the visual data, keeping only the specific details that prove who you are. The challenge for scientists is to figure out which method of shrinking the image keeps the identity intact while discarding the rest. If the compression is too harsh or uses the wrong technique, the machine might fail to recognize the person, or worse, mistake one person for another.
A team of researchers set out to solve this puzzle by testing how well different compression tools preserve a person's identity under these extreme conditions. They did not just look at how the images looked to the human eye; they measured whether the compressed images still worked for the computer systems that verify identity. They tested ten different standard compression methods, ranging from old, familiar formats to new, advanced ones, and also built their own custom tool designed specifically for this task. They ran these tests on thousands of face images, compressing them down to 1,024 bytes and even pushing the limit to just 512 bytes. They then fed the shrunken images into powerful face-recognition software to see if the system could still tell the difference between a real match and a stranger.
The results revealed a clear split between the old ways of doing things and the new. When the researchers compressed the images to 1,024 bytes, modern tools like WebP, AVIF, and a new standard called JPEG-AI performed remarkably well. They kept the identity information so intact that the computer made very few mistakes, performing almost as well as it would with a full, uncompressed photo. However, older methods like the classic JPEG format and JPEG 2000 struggled significantly. While they could fit the image into the small space, they distorted the features so much that the computer often failed to recognize the person. The situation became even more critical when the researchers tried to squeeze the images down to just 512 bytes. At this extreme limit, most of the modern tools began to fail, with error rates skyrocketing. Only a few specialized methods, including the researchers' own custom-built compressor and the JPEG-AI standard, managed to hold up, keeping the identity recognizable even in this tiny space.
One of the most surprising discoveries was that a picture that looks perfect to a human does not necessarily work well for a computer. The researchers found that some compression methods produced images that looked very sharp and clear to the eye but actually scrambled the specific details the computer needed to identify a person. Conversely, some images that looked slightly blurry or imperfect to a human were actually the best for the computer to use. This means that judging a compression method by how pretty the picture looks is a poor way to decide if it will work for security. The researchers also looked at whether these compression methods treated people of different backgrounds fairly. They found that while compression slightly increased the difficulty of recognizing certain groups, one older method, JPEG 2000, made the problem much worse, amplifying the gap in recognition accuracy between different skin tones. The modern methods, including the custom tool, kept this unfairness to a minimum.
The study also examined what happens when an image is compressed, then compressed again, or when someone tries to trick the system with a hidden digital attack. They found that re-compressing an image with certain modern tools was safe, but mixing specific older and newer formats could destroy the identity data entirely. Furthermore, the act of compressing an image to such a small size naturally removed some types of digital tricks that attackers use to fool face scanners, acting as an accidental shield. The researchers concluded that for systems that need to store a face in a tiny space, the choice of compression tool is critical. They recommend using modern, specialized formats like JPEG-AI or WebP, and avoiding older standards that fail at these tight limits. Their custom tool proved to be a strong contender, especially when the space was extremely limited, offering a way to keep digital identities secure even when the storage budget is almost zero.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.