Defence Science and Technology Group, PO Box 1500, Edinburgh, SA 5111, Australia.
Forensic Sci Int. 2024 Nov;364:112202. doi: 10.1016/j.forsciint.2024.112202. Epub 2024 Aug 23.
Facial recognition plays a vital role in several security and law enforcement workflows, such as passport control and criminal investigations. The identification process typically involves a facial recognition system comparing an image against a large database of faces to return a list of probable matches, called a candidate list, for review. A human then looks at the returned images to determine whether there is a match. Most evaluations of these systems tend to examine the performance of the algorithm or human in isolation, not accounting for the interaction that occurs in operational contexts. To ensure optimal whole system performance, it is important to understand how the output produced by an algorithm can impact human performance. Anecdotal claims have been made by users of facial recognition systems that the images being returned by new algorithms in these systems have become more similar in appearance compared to old algorithms, making their job of determining the presence of a match more difficult. This paper explores whether these claims are true and whether the latest facial recognition algorithms decrease human performance compared to an old algorithm from the same company. We examined the performance of 40 novice participants on 120 face matching trials. Each trial required the participant to compare a face image against a candidate list containing eight possible matches returned by either a new or old algorithm (60 trials of each). Overall, participants were more likely to make errors when presented with a candidate list from a new algorithm. Specifically, they were more likely to misidentify an incorrect identity as a match. Participants were more accurate, confident, and faster on candidate lists from the older algorithm. These findings suggest that new algorithms are generating more plausible matches, making the task of determining a match harder for humans. We propose strategies to potentially improve performance and recommendations for future research.
人脸识别在许多安全和执法工作流程中起着至关重要的作用，例如护照管制和犯罪调查。识别过程通常涉及人脸识别系统将图像与大型面部数据库进行比较，以返回一个可能的匹配列表，称为候选列表，供审查。然后，人类会查看返回的图像以确定是否存在匹配。这些系统的大多数评估往往只检查算法或人类的性能，而不考虑在操作环境中发生的交互。为了确保整个系统的最佳性能，了解算法产生的输出如何影响人类的性能非常重要。人脸识别系统的用户曾有过这样的说法，即这些系统中新算法返回的图像与旧算法相比，外观变得更加相似，使得他们确定匹配的存在变得更加困难。本文探讨了这些说法是否属实，以及最新的人脸识别算法是否会降低与同一家公司的旧算法相比人类的性能。我们检查了 40 名新手参与者在 120 次面部匹配试验中的表现。每个试验要求参与者将一张面部图像与候选列表进行比较，候选列表包含新算法或旧算法返回的 8 个可能的匹配项（每个算法 60 次试验）。总的来说，当参与者看到新算法生成的候选列表时，他们更容易出错。具体来说，他们更有可能将错误的身份误识别为匹配。参与者在旧算法的候选列表上更准确、更自信、更快。这些发现表明，新算法生成的匹配项更合理，使得人类更难以确定匹配。我们提出了潜在的提高性能的策略和对未来研究的建议。
