How AI Is Changing Support for Victims of Sexual Image Abuse
- May 20
- 6 min read
At PAPS, two engineers are currently developing an AI-powered system to support people who have suffered sexual image abuse.
Once an image spreads on the internet, finding it again is extremely difficult. Images end up in many places—social media, message boards, overseas sites, video sites—and there is a limit to how long humans alone can keep searching.
Since 2019, with a grant from the Mitsubishi Foundation, PAPS has continued research and development on countermeasures using machine learning. Since fiscal 2022, we have also been working in earnest on developing an AI-based support system. With AI technology advancing dramatically over the past year, new possibilities are opening up for work that previously had to rely on human hands.
ProtectionAI: a system for finding images of victims
ProtectionAI, currently under development, is a system for searching whether a victim's sexual images have spread across the internet and social media.
Based on images provided by the victim, the AI reads facial features and converts them into a sequence of 512 numbers—known in technical terms as a "feature vector". A facial photo cannot be reconstructed from a feature vector, so registration is done in a privacy-conscious form.
The system then compares these against images collected from the internet and judges that they "may be images of the same person". For this comparison we use "cosine similarity", a computational method for measuring how similar two pieces of data are.
ProtectionAI currently trains and runs analysis on a dataset of roughly 100 million images.
We combine and validate multiple approaches while drawing on publicly available AI technology. At the same time, out of consideration for preventing misuse and ensuring safety, we do not disclose the detailed model architecture or how the system is operated.
Paprika: the system that goes searching across the internet
Finding images of victims requires going out to search many corners of the internet. For this, PAPS is developing its own crawler system, "Paprika". A crawler is a program that automatically browses websites and collects the information it needs. The name "Paprika" is a coinage combining "PAP" from PAPS with "rika".
Images of victims spread to many different places—social media, overseas sites, anonymous message boards, video sites. What's more, each site stores and surfaces images differently, which makes ordinary crawlers a poor fit for the task.
It is published here:https://paps-jp.github.io/paprika/
Previously, we also explored automating operations by reading the DOM, a web page's internal structure. The DOM comprises the buttons, input fields, images, links, and other elements inside a page. But real-world websites are complex and each is built differently, so this approach often failed.
So we now use Vision AI—AI that "looks at and understands the screen". Just as a person looks at a browser window and judges "here is an input field" or "this button looks right", the AI works out how to operate the page by looking at it.
Paprika also incorporates a mechanism in which the AI automatically writes and tests code for retrieving images inside a safe experimental environment. This safe environment is called a "sandbox"—an isolated space where failures cannot affect outside systems.
In AI development, "breaking things into small pieces" matters
AI development is not just about using high-performance AI; how the overall system is built matters too.
Rather than bundling AI features into one large system, PAPS emphasizes "micro-APIs": splitting functionality into small pieces used separately.
An API is like an entry point for calling a piece of functionality from outside. For example, we keep functions small and separate: reading images, extracting facial features, searching for similar images, operating takedown-request forms, and so on.
Large AI systems—especially processes that use GPUs—can take a long time to start up. A GPU is a device for performing AI computations at high speed. By splitting things so that only the necessary functions run when needed, we process work efficiently and can locate the cause more easily when problems occur.
Toward developing our own AI model
ProtectionAI reads facial features from images provided by victims themselves and matches them against images on the internet.
At present, the core of this face recognition relies on a high-performance face recognition AI model. It is highly accurate and plays a key role in finding images of victims, but in the future we aim to develop and operate PAPS's own face recognition AI model.
The reason is that supporting victims of sexual image abuse means handling extremely sensitive information: victims' faces. It is therefore essential to be able to strictly control—in ways suited to frontline support—what data is used, how the model is trained, how matching is performed, and who has access.
With our own model, we could build in accuracy improvements specialized for victim support, safety-oriented design, and anti-misuse mechanisms in much finer detail.
On the other hand, developing our own face recognition AI model requires substantial computing power, which makes computational resources such as GPUs indispensable. A GPU is a device for performing AI computations at high speed.
PAPS currently operates two datacenter-class servers with AMD EPYC processors, three PC servers, three NAS units, and seven GPUs. AMD EPYC is a high-performance server CPU designed for heavy workloads, and a NAS is a high-capacity storage device accessible from multiple servers.
We were also able to secure 32TB of SSD storage before prices surged, and are building out an environment for handling big data.
The battle with heat and noise
In AI development, "heat" and "noise" are major challenges alongside computing power.
GPUs generate a great deal of heat when performing AI computations. Under sustained heavy load, temperatures climb too high, degrading performance and shortening the hardware's lifespan. Keeping GPU temperatures around 60°C is therefore important.
At PAPS, we attach heat sinks to the GPUs and use multiple circulators and fans to keep heat from building up. A heat sink is a metal component that dissipates a machine's heat. Last summer was brutal, so we need to plan more serious heat countermeasures before this summer arrives.
Because the servers and GPUs run inside our office, cooling fan noise is also an issue. We therefore use many large 12cm fans, which can move enough air even at low rotation speeds.
AI development is not only about algorithms. It is also a battle with very down-to-earth problems: heat, noise, electricity, and physical space.
The frightening side of "search technology"
At the same time, face-search technology carries serious challenges. When we began development, we never imagined it would become able to find people's faces with such accuracy. In actual operation, there were moments when it found images with greater precision than we had assumed possible—we were astonished by the power of big data and AI, and frightened by it as well. It was even able to find images of a person harmed nearly ten years ago that no longer surface in web searches.
Being able to search for victims' images also means, conversely, that whether someone suffered sexual image abuse in the past could potentially be looked up in a matter of seconds.
Technology built to help victims could be used to surveil someone, expose their past, or enable discrimination and harassment. That is precisely why PAPS places such importance on building mechanisms that cannot be misused.
There is also the difficulty that, in order to search for images of victims, those providing support must view the victim's face. "I want it found"—and at the same time, "I don't want it seen". Support for sexual image abuse carries this dilemma. PAPS moves development forward while weighing both the support society needs and how to protect each individual's privacy.
Advances in AI technology are expanding the possibility of automating parts of the work that once demanded long hours of human labor. But in supporting victims of sexual image abuse, AI is no magic wand. There are challenges of accuracy, cost, heat and noise, security—and of protecting victims' privacy and dignity.
PAPS is not developing AI because we want to use new technology. The purpose is to help those who have been harmed reclaim their lives as quickly and safely as possible.
At the same time, facial recognition AI, if used wrongly, carries the danger of being misused for surveillance or harassment. That is why PAPS places great weight on rules governing who may use it, for what purpose, and to what extent.
To prevent misuse, there is also information that should not be made public, such as API specifications and detailed search methods. While valuing transparency, we carefully separate the information that can be disclosed from what must be withheld to keep victims safe.
Privacy protection is our most important principle. At the same time, finding images spread against a person's will and connecting them to takedown requests serves a strong public interest. What matters is using the technology under strict controls, to the minimum extent necessary, and premised on the person's consent.
PAPS continues development—facing squarely both the convenience and the dangers of AI—to protect the dignity of those who have been harmed.
