SENTINEL: Securing Age Estimation and Digital IDs against Presentation and Injection Threats

随着全球立法海啸般要求互联网具备“年龄意识”，对在线年龄保障的需求将继续呈指数级增长。基于人工智能的年龄估计，从生物特征或行为特征中识别年龄，是年龄保证的关键手段，特别是对于没有文件或记录的儿童和其他弱势群体。但这些方法越来越容易受到人工智能产生的攻击，以及对它们的脆弱性的感知，这可能会让用户在网上伪造年龄，严重破坏对年龄保证技术的信心。众所周知，生物识别系统容易受到表示/欺骗攻击（PADs）和注入攻击（iad）的攻击。然而，在过去的12个月里，消费者设备和应用程序上越来越复杂的系统已经开始让任何人更容易对系统进行基本的人工智能生成的面部（deepfake）和音频欺骗攻击。事实上，2023年在网上发现了80多个深度假冒应用程序，而且它们的质量一直在提高。以前的研究是由我们的瑞士研究伙伴IDIAP进行的，研究各种攻击向量和合成数据，以及由AVID进行的测试标准化。我们希望在这项研究的基础上，进一步提高对这些攻击的防御能力，不仅是为了我们的瑞士年龄验证合作伙伴private SA，也是为了通过他们的行业协会AVPA与整个AV行业分享经验教训。该项目将开发识别人工智能生成的演示和注入攻击的技术，这些攻击涉及新兴的年龄保证方法，包括面部分析、声纹分析和游戏。然后，它将标准化方法来测试反病毒提供商针对它们实施的防御措施。AVID全面的测试程序和最先进的技术将保护整个行业的诚信客户，使其免受未成年用户的不当访问，以及与之相关的法律、商业和道德后果。与此同时，AVPA与业内其他供应商协商，以确保对这种威胁做出联合反应，这种威胁可能会破坏各种形式的在线年龄保证的可信度。这项新技术将有助于为儿童提供增强的保护，使其免受可能对他们或他人造成伤害的商品、内容或服务的侵害，并且恰逢全球监管变化要求有效和准确的年龄保证系统来保护儿童免受伤害。

With a tsunami of global legislation demanding that the Internet becomes "age-aware", the demand for online age assurance will continue to grow exponentially. AI-based age estimation, with age discerned from biometric or behavioural features, is a critical means of age assurance, particularly for children and other vulnerable groups without documentation or records. But these methods are increasingly becoming vulnerable to AI-generated attacks, and the perception of vulnerability to them, which could allow users to fake their age online and seriously undermine faith in age assurance technologies.It is well established that biometric systems are vulnerable to presentation/spoofing attacks (PADs) and injection attacks (IADs). However in the last 12 months, increasingly sophisticated systems available on consumer devices and applications have begun to make it easier for anyone to conduct basic AI generated facial (deepfake) and audio spoof attacks on systems. Indeed, there were over 80 deep fake apps found online in 2023 and they are observed to be improving in quality all the time.Previous research has been conducted by our Swiss research partner, IDIAP, on various attack vectors and synthetic data, and by AVID on the standardization of testing. We want to build on this research and advance it to create defences against these attacks, not only for our Swiss Age Verification partner, Privately SA, but also to share the general lessons learnt with the AV industry as a whole through their trade association, the AVPA.This project will develop technology identifying AI-generated presentation and injection attacks across emerging methods of age assurance, including facial analysis, voiceprint analysis and game play. It will then standardize methods to test the defences AV Providers implement against them.AVID's comprehensive testing procedures and state-of-the-art technology will protect the integrity clients of the industry as a whole from allowing inappropriate access to underage users and the legal,commercial and ethical consequences associated with that, while the AVPA consults with other providers in the industry to ensure a joined up response to this threat which could undermine credibility in all forms of online age assuranceThis new technology will help provide children enhanced protection from exposure to goods, content or services that may cause them or others harm, and comes at the perfect time as global regulatory changes mandate effective and accurate age assurance systems to protect children from harm.