Today, my FotoForensics service turns 12 years old! It has over 6,800,000 unique pictures. That’s up about 800,000 from last year. The site’s popularity has increased by about 12%, from being in the top 80,000 of internet destinations last year to about 70,000 right now. (By comparison, the top five internet destinations are currently Google, Facebook, Youtube, [surprisingly] Twitter, and Instagram. None of my services are that popular, yet.)
I actually find the site’s popularity kind of surprising, considering all of the wide-area bans. For example:
I’ve been actively blocking uploads from the Tor network since 2017 (7 years and counting). This is because over 90% of uploads from Tor violated my terms of service. Besides the IDs (for fraud) and bulk uploading, Tor users account for over 50% of the uploaded child porn (yuck). Initially, the bans were almost set manually, but I quickly added in automated rules. Today, the detection of Tor uploads is completely automated. While I see the “Tor blocked” messages in the logs, I don’t give it a second thought.
Before being banned, Russia was almost abusive as Tor. Bulk uploading, porn, child porn, fake IDs, drug distribution photos, and much more. It wasn’t any single user or group; it was wide-spread and endemic. When I’d catch them, I’d ban them. However, the Russian mentality when getting caught is to evade the ban and then increase in volume. By 2020, I’d had enough of this and issued a country-wide ban.
To put this ban into perspective, the public FotoForensics site does not permit uploads of ID, drivers licenses, passports, etc. I’ve been keeping metrics on these types of uploads:
Black and white means I’ve never received any ID pictures from that country. (E.g., You’ll never see a Greenland passport because it’s part of Denmark.) The volume per country goes from bright green (few sightings) to bright red (the most sightings). Even with zero uploads from Russia for the last 4 years, Russia still has more ID uploads than any other country. Russia was so prolific that four year after banning their uploads, Russia still has uploaded more IDs than the next three countries combined (China, Syria, and Indonesia).
Since putting the country-wide ban in place, I’ve received many unban requests from Russians, including:
People claiming to be Russian law enforcement. As proof of their legitimacy, some provided photos of fake police IDs and fake police badges. Others gave fake names and addresses for police stations that don’t exist. None of these requests appear to be real law enforcement.
A few Russian service providers have asked to be unbanned. However, they were part of the original problem and have done nothing to mitigate the abuse that emanates from their networks. One was rather chatty and made it clear: they should not be held responsible for abuses from their customers, even if the problem comes from a majority of them.
Two people claiming to be with Russian media asked to be unbanned. However, I couldn’t authenticate either of them and they were coming from shared networks where people (let’s assume other people and not them) were uploading prohibited content.
When banned, these users would occasionally switch to VPNs or proxy networks and continue their uploads. These abuses would be quickly caught and all uploads from the proxy network would be blocked. Today, it is very difficult for anyone in Russia to upload anything to FotoForensics. And if they find a way to upload and violate my terms of service (porn, IDs, etc.), then they get blocked fast.
Because of the country-wide ban, I’ve also received a lot of hate email from Russians. They call me a tool of the West, a ‘liberal’, or pro-Ukrainian propagandist. My thinking here is pretty simple: their country was banned for widespread abuses. Writing hostile messages to me don’t alleviate the abuse.
I really went out of my way to not ban all of China, but eventually I had no other option.
The first country-wide ban lasted a few days. The ban message even explicitly mentioned why there was a country-wide ban and included sample images. When the ban lifted, the abuses resumed a few days later.
The second country-wide ban lasted a week. Again, it resumed when the ban lifted.
The third country-wide ban lasted a month. But, again, it didn’t stop the problem.
China has been banned for a few months now, and the ban will last for one year.
Because Russians managed to get most free or inexpensive VPNs and proxy services banned, it’s really hard for Chinese people to evade the ban.
Unlike Russia, the abuses from China are not from a majority of Chinese citizens. It seems to come from a handful of very prolific groups. However, because of how China has allocated their national internet service, I have no easy method to distinguish the prolific abusers from the non-abusive population. I had expected to see a drop in volume at FotoForensics when I banned China. Surprisingly, this didn’t happen. Instead, I saw a 5% increase in uploads and these new uploads were not violating my terms of service! I suspect that people in China are talking to people outside of China, and that is driving more volume to my site.
You can’t spell “pain” without “ai”
The public FotoForensics service is explicitly used for research. (It’s mentioned 18 times in the FAQ.) When doing any kind of research, you need sample data. Most researchers go out and harvest content. This creates a bias due to their search criteria. With FotoForensics, I just wait for people to upload real-world examples. Of course, this does introduce a different bias: the service receives pictures when people question the legitimacy or are testing; it rarely receives unquestioned legitimate pictures.
This bias allows me to determine what people are most concerned about. And right now? It’s AI. Whether the picture is completely AI generated or altered using AI systems, people have legitimate concerns. Here are some recent examples (click on the picture to view at FotoForensics):
This first picture appears to show a child soldier in Ukraine. However, the metadata was stripped out; we cannot identify the source. (Also, you might not notice that his limbs look unnatural.) In this case, Stability AI (the company behind Stable Diffusion) has been adding watermarks to some of their pictures. These show up as binary repeating dot sequences in the compression level (ELA):
I can easily detect this watermarking; this picture came from Stable Diffusion before the metadata was removed. However, I haven’t figured out how to decode the watermark yet. (Help would be appreciated!) I wouldn’t be surprised if it encoded a timestamp and/or account name.
Not every AI-generated image is as easy to detect as Stable Diffusion. Consider this picture, titled “Mea Sharim”, which refers to an ultra-Orthodox Jewish neighborhood in Jerusalem:
At first glance, it appears to show an Orthodox man standing on a corner. However, the ELA identifies an AI-generated erasure:
The erased area is adjacent to the man and appears a little taller. The erased area also seems to have an erased shadow. With some of these “smart erase” tools, you can select an area and have the AI fill in the void. That appears to be the case here. While I can conclusively detect the altered area, I can only speculate about what was erased. I suspect that the visible man was facing another person. The artist selected the other person and his/her shadow and performed a smart erase.
Both of these pictures could easily be used for propaganda. Having metadata that identifies a vague statement about a digital alteration using AI means nothing when the metadata can be trivially removed and replaced.
Usually I try to celebrate each new year at FotoForensics with a new feature release. For the last few months, I’ve been doing a lot of behind-the-scenes modifications. This year, rather than doing one big update, I plan to continue the rolling updates with more changes coming down the line. This includes addressing some of the AI-generated content later this year. (If you’re on the commercial service, then you should be seeing some of the new detection systems right now in the metadata analyzer.)
After 12 years, I’m still making regular improvements to FotoForensics, and enjoying every minute of it. I’m very thankful to my friends, partners, various collaborators, and the public for 12 years of helpful feedback, assistance, and insights. This year, I especially want to thank my mental support group (including “I’m Bill not Bob”, “I’m Bob not Bill”, and Dave), my totally technical support group (Marc, Jim, Richard, LNM, Troy’s evil cat, and everyone else), Joe, Joe, Joe, AXT, the Masters and their wandering slaves, Evil Neal, Loris, and The Boss. Their advice, support, assistance, and feedback has been invaluable. And most importantly, I want to thank the literally millions of people who have used FotoForensics and helped make it what it is today.