Basic visual manipulations can fool some FAE, but there’s a solution, says study

Facial age estimation (FAE) has emerged as a popular option for websites covered by age assurance regulations, offering a biometric method that does not rely on reference to a government ID or the collection of sensitive personal information.

However, concerns have arisen that some underage users are able to easily fool FAE systems with simple manipulations, such as drawing a fake moustache on their face; a recent report from Internet Matters shows that 46 percent of UK children think that age checks mandated by the Online Safety Act are easy to bypass, and a third (32 percent) of children have bypassed age checks by entering a fake birthday, drawing on facial hair, or other easy workarounds.

A new paper addresses these gaps in the safety net. The work, a collaboration by three Greek academics, presents “a systematic study of age verification robustness by simulating visual alterations that can be easily achieved by underage individuals.” Testing spanned seven AI FAE models, three data sets and four types of manipulation – adding beard stubble, a moustache, eye makeup, and lipstick.

The focus was on “spurious correlations present within the training data” – attributes that are “predictive of the target label in the training data,” but “not causally tied to the target concept.”

In face analysis, such attributes include hair color, makeup, cosmetics, etc. These attributes can be easier to learn than the actual target attribute, and when they co-occur strongly with the target labels, models tend to rely on them to make decisions. Thus, presented with a six-year-old wearing lipstick, the model may default to “lipstick equals adult,” and misclassify the subject as a result.

Bias mitigation can help adjust results

The results of the study will surely be logged by critics of age assurance technology as proof it doesn’t work as it’s supposed to. “Despite reasonable performance on the original images,” the authors conclude, “most models show manipulation-specific vulnerabilities.”

Synthesized beard stubble is “the most damaging manipulation, producing an average flip rate of 20.8 percent from true negatives to false positives.”

Vulnerabilities vary across demographic groups, with female subjects “generally more affected than male subjects” and people of Indian descent more affected by beard stubble manipulations – more fuel for those who argue that algorithmic bias makes biometric age assurance tools less fair to some groups than to others.

The research, however, also points to a viable solution – namely, bias mitigation methodologies in lightweight linear probe settings. “Specifically, we train a lightweight two-layer MLP classifier on the frozen extracted features to evaluate whether performance can be recovered under manipulation-induced shifts.” Overall, “results suggest that robust age verification under simple appearance manipulations can be improved through lightweight adaptation on top of the pretrained backbones.”

Like many biometric services, age assurance is becoming a layered affair: single, isolated approaches are being subsumed into larger service models that aim to solve the core problem one way or another, rather than peddling one specific method. So the researchers’ caveat is significant: “our results pertain to the robustness of the visual component specifically; whether these manipulation-induced errors survive as bypasses within a full multi-signal system, where other signals could compensate for an incorrect visual estimate, remains an open question.”

Facial age estimation, as such, “should remain a component of a broader verification process.”

Support from kids will carry age checks into future

Ultimately, the effectiveness of one biometric method for age assurance matters less than the project at large. While Internet Matters’ study expresses concern that kids are still encountering harmful material online, it also notes evidence of progress: “new laws are making safety measures more prominent across children’s online spaces, with parents and children largely welcoming these changes.”

Indeed, the voices that matter most in this issue are those belonging to kids – and they largely seem to support efforts to curb the amount of extreme content they consume. “Age checks easy to circumvent” is a more sensational headline, but the findings also show that, for kids, “age checks are common and seen as easy to complete.”

“About half of children (53 percent) say they have recently been asked to verify their age, most often when setting up new accounts. According to children, the following methods were described as easy: uploading a government ID document (88 percent), facial age estimation (89 percent) and using a third-party app (88 percent).

The secret ingredient for any major social change is time. As age checks become more commonplace, they are likely to become less controversial: people will get used to them as a normal feature of online life, and accumulated evidence will build the case for continuing down a particular path, or pivoting to ensure efforts are more effective. Buried in the data is the finding that “parents and children are optimistic the online world is getting safer for children.” There is hope that things can be better – and that will only drive the effort to make it so.

Article Topics

age verification  |  biometric age estimation  |  biometrics  |  biometrics research  |  facial age estimation (FAE)