“You don’t take a photograph, you make it.”
Ansel Adams, Photographer, 1902-1984.
Who would you prefer to interact with online? A person whose avatar features a photo of their smiling face, and whose social media includes scores of images of them mixing with friends and family across a variety of different locations? Or a person whose avatar is the generic site default, and whose social media features only a scant handful of photos, none of which include their face?
It used to be the case that to establish if a person online was genuine, and not an automated bot or a false persona controlled by one or more other individuals, one needed only to check their social media images. Potential red-flags included:
- Use of a default system avatar.
- An avatar that does not feature the person’s face, such as a logo or cartoon.
- An absence of personal imagery across their social media platforms.
- Images where the person’s face is obscured or turned away from the camera.
- No photos of the person in social settings.
- No photos of the person in different locations, such as on holiday or visiting public sites.
Even if the person’s imagery did include their face, additional checks might prove prudent. For example, one could conduct a reverse image search to check that their facial photos had not been sourced from stock imagery, including images that had been mirror-flipped or otherwise manipulated (such as changing the colour of clothing, eye colour, etc.). This practice was especially common in dating fraud, and the use of stock images led to many early online scammers being busted (e.g. Chan, 2012; Sackville, 2016).
For many years it has proven extremely challenging to construct or to generate synthetic facial images that ‘feel’ authentic when viewed by a casual human observer. Traditional approaches for producing false faces included blending together images of two genuine faces to create a new third face – a time-consuming and exacting process because of the number of resultant artefacts requiring integration and smoothing by hand. Critically, when human observers view the resultant images, they often develop an immediate sense that something is ‘off’, resulting in feelings of discomfort or suspicion. Paying attention to these feelings is fundamental to the detection of all deception, as they result from our subconscious processes of pattern recognition and detection of anomalous features that differ from our experience.
Of course, an absence of facial imagery is no guarantee that a person’s profile is false, nor is it the only indicator. However, profiles that lack facial photographs suggest that either their owner has either intentionally provided the minimum details necessary to open an account and access a site; that the individual values their privacy; or that the account is false and has been created to enable some form of online manipulation.
The challenges involved in generating convincing facial images have decreased significantly over the past couple of years, and facial photographs no longer indicate credibility. The capability now exists to automate the generation of credible synthetic imagery of any object, scene or person, creating a convincing photographic record of people and things that have never existed (Nguyen et al., 2015; Sharif et al., 2016; Karras et al., 2017).
A Generative Adversarial Network (GAN) is an algorithm that uses a pair of neural networks that compete to learn the statistical characteristics of a training dataset. Once the algorithm has discerned a dataset’s essential features, it can then generate new data with similar features. The graphics processing company NVIDIA has pioneered the use of GANs to learn and create convincing images of people’s faces, rooms, buildings, animals, vehicles, phones, etc. (Karras, Aila, et al., 2018; Karras, Laine, et al., 2018; Karras, Laine, et al., 2020). To see examples of the synthetic faces produced by their GAN, go to thispersondoesnotexist.com (note that refreshing the page will render a new face). Other applications, including the generation of artificial landscapes, mapping one animal’s characteristics onto another’s, and simulation of dynamic environments with no underlying behaviour engine, can be seen here.
The faces (currently) produced by GANs are not always perfect. Indicators of falsehood are sometimes visible in photographic anomalies such as a person’s teeth looking blurry or superimposed over their mouth, their hair blending into the background at its ends, and the tips of the person’s ears appearing blurry, distorted or missing. GANs can struggle to render realistic earrings, glasses, hats and headscarves. Other faces in the background of a GAN-generated image may feature significant distortion, and sometimes the background itself appears odd or unnatural. However, most of the faces generated by these systems are highly convincing. And so far as a scammer is concerned, if the system generates an image that is not convincing, they simply move on to the next image that is.
The practice of using GAN imagery to bolster false online personas is proliferating rapidly. In 2020 several conservative outlets (including the Washington Examiner, RealClear Markets, American Thinker, and The National Interest) published stories about the Middle East that were critical of Qatar and promoted stricter sanctions on Iran (Kirell, 2020; Vincent, 2020). The articles were authored by a network of at least 19 fake personas that over a year placed more than 90 opinion pieces in 46 different publications, all supporting a broader Middle East propaganda campaign. The fictitious authors’ online profiles featured both mirror-flipped facial imagery of real journalists, together with synthetic faces lifted directly from thispersondoesnotexist.com.
Increasingly, GANs are now being used to generate synthetic video footage. Training footage is first analysed to acquire its inherent and emergent dynamic visual properties. Artificially rendered footage then reflects these characteristics and appears to be a real video stream. For example, NVIDIA (once again) has developed a driving simulator that presents drivers with video footage of a street scene that responds dynamically to their driving, featuring other cars, buildings, trees, pedestrians, etc. (Vincent, 2018; Wang et al., 2018). Footage of the GAN-generated video can be seen here.
These kinds of technologies used to be the sole preserve of big-budget intelligence agencies. However, they are now open to the public for free as researchers give away their source code and training data sets. The capabilities of GANs will only improve exponentially as computing power increases in parallel with advances in machine learning, enabling the generation of exceptionally realistic, compelling, and credible images and video.
Given that humans now struggle to distinguish synthetic faces from genuine faces, how might such falsehood be detected? One approach suggests that we may be able to train observers to detect synthetic faces. Work by Robertson et al. (2018) has found that observers’ initial differentiation between morphed (blended) facial images and real faces corresponds with chance (48%). However, after observers received training to detect the anomalies that may be present in false photos, the level of correct differentiations rose to 89%. It is worth noting, however, that this study only assessed post-training detection of blended imagery, and that such human differentiation may be challenged significantly as the quality of imagery generated by GANs continues to improve.
Another approach is the application of machine learning to detect false faces. In principle, if machine learning can generate a synthetic (yet convincing) face, then machine learning should also be able to detect a synthetic face. To do so requires training GANs to learn, and to recognise the unique features of false faces that humans struggle to detect. Many initiatives are underway to develop such systems (e.g. Khodabakhsh et al., 2018; Liu, 2018; Natsume et al., 2018; Rossler et al., 2019), with one study already claiming 94% accuracy in the automatic detection of GAN-generated faces (Tariq et al., 2018). An extensive research programme in this area is being pursued by the US Air Force, which is funding a $15M/year project with the Massachusetts Institute of Technology to explore deception in Artificial Intelligence and image recognition (Reim, 2020). The US Defense Advanced Research Projects Agency is similarly investigating these issues under its Techniques for Machine Vision Disruption project, with the goal to “develop specific techniques to disrupt neural net-based computer vision technology in situations where neither the original training set nor the specific vision architecture is available.” (Defense Advanced Research Projects Agency, 2020).
From this point forward, pictures may now paint countless lies.
References
Chan, C. (2012). Did Chick-fil-A Pretend to Be a Teenage Girl on Facebook? https://gizmodo.com/did-chick-fil-a-pretend-to-be-a-teenage-girl-on-faceboo-5928926.
Defense Advanced Research Projects Agency. (2020). Artificial Intelligence Exploration (AIE) Opportunity DARPA-PA-19-03-06: Techniques for Machine Vision Disruption (TMVD). Retrieved 10/09/2020 from https://www.defencescienceinstitute.com/images/TMVD_AIE_DARPA-PA-19-03-06.pdf.
Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017). Progressive growing of GANS for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196.
Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2018). Progressive Growing of GANS [Generative Adversarial Networks] for Improved Quality, Stability, and Variation. Paper presented at the International Conference on Learning Representations 2018, Vancouver, BC.
Karras, T., Laine, S., & Aila, T. (2018). A Style-Based Generator Architecture for Generative Adversarial Networks. arXiv, 1812.04948v1.
Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., & Aila, T. (2020). Analysing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8110-8119.
Khodabakhsh, A., Ramachandra, R., Raja, K., Wasnik, P., & Busch, C. (2018). Fake Face Detection Methods: Can They Be Generalized? Paper presented at the 2018 International Conference of the Biometrics Special Interest Group (BIOSIG).
Kirell, A. (2020). Right-Wing Media Outlets Duped by a Middle East Propaganda Campaign. Retrieved 09/09/2020 from https://www.thedailybeast.com/right-wing-media-outlets-duped-by-a-middle-east-propaganda-campaign.
Liu, H., Zheng, S., Hao, S., & Li, Y. (2018). Multifeature Fusion Detection Method for Fake Face Attack in Identity Authentication. Advances in Multimedia, 2018, 1-10.
Natsume, R., Inoue, K., Fukuhara, Y., Yamamoto, S., Morishima, S., & Kataoka, H. (2018). Understanding Fake Faces. Paper presented at the European Conference on Computer Vision (ECCV), Munich, Germany.
Nguyen, A., Yosinski, J., & Clune, J. (2015). Deep neural networks are easily fooled: High confidence predictions for unrecognisable images. Paper presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
Reim, G. (2020). US Air Force grapples with vexing problem of AI spoofing. Retrieved 10/09/2020 from https://www.flightglobal.com/defence/us-air-force-grapples-with-vexing-problem-of-ai-spoofing/139973.article.
Robertson, D. J., Mungall, A., Watson, D. G., Wade, K. A., Nightingale, S. J., & Butler, S. (2018). Detecting morphed passport photos: a training and individual differences approach. Cognitive Research: Principles and Implications, 3, 27.
Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., & Nießner, M. (2019). Faceforensics++: Learning to detect manipulated facial images. Paper presented at the Proceedings of the IEEE International Conference on Computer Vision.
Sackville, K. (2016). This guy would be the perfect boyfriend except for one thing. Retrieved 09/09/2020 from https://nypost.com/2016/03/10/i-got-catfished-by-a-stock-image/
Sharif, M., Bhagavatula, S., Bauer, L., & Reiter, M. K. (2016). Accessorise to a crime: Real and stealthy attacks on state-of-the-art face recognition. Paper presented at the 2016 ACM SIGSAC Conference on Computer and Communications Security.
Tariq, S., Lee, S., Kim, H., Shin, Y., & Woo, S. S. (2018). Detecting Both Machine and Human Created Fake Face Images In the Wild. Paper presented at the Proceedings of the 2nd International Workshop on Multimedia Privacy and Security, Toronto, Canada.
Vincent, J. (2018). Nvidia has created the first video game demo using AI-generated graphics. Retrieved 11/12/2018 from https://www.theverge.com/2018/12/3/18121198/ai-generated-video-game-graphics-nvidia-driving-demo-neurips.
Vincent, J. (2020). An online propaganda campaign used AI-generated headshots to create fake journalists. Retrieved 09/09/2020 from https://www.theverge.com/2020/7/7/21315861/ai-generated-headshots-profile-pictures-fake-journalists-daily-beast-investigation.
Wang, T.-C., Liu, M.-Y., Zhu, J.-Y., Liu, G., Tao, A., Kautz, J., & Catanzaro, B. (2018). Video-to-Video Synthesis. Cornell University Open Access. Retrieved 14/12/2018 from https://arxiv.org/pdf/1808.06601.pdf.