1 of 45

Computer Vision, Society, and Ethics

CS5670: Computer Vision

Closed circuit TV monitoring at the Central Police Control Station, Munich, 1973

(from the Wikipedia article on the Panopticon)

2 of 45

Announcements

  • In class final Tuesday, May 6
    • 2 sheets of notes (front and back) allowed
    • Final is comprehensive (covers entire course)

​

  • Project 5, Part A due this Friday, April 25, by 8pm
  • Project 5, Part B due Monday, May 5, by 1pm

3 of 45

Additional Resources

  • FATE (Fairness Accountability Transparency and Ethics) in Computer Vision Tutorial

​

  • https://exposing.ai/
    • Adam Harvey and Jules LaPlace

​

  • Foundations of Computer Vision, Chapter 4

4 of 45

Advances in computer vision

  • Sometimes we think of technological development as a uniform positive
  • But computer vision exists in a societal context, and can have both good and bad consequences – need to be mindful of both
  • Example: as computer vision gets better, our privacy gets worse (e.g., through improved face recognition)

5 of 45

Today

  • Examples of bias in computer vision and beyond
  • Datasets and unintended consequences
  • DeepFakes and image synthesis methods

​

6 of 45

Questions to ask about a specific task

  • Should I be working on this problem at all?
  • Does a given vision task even make sense?
  • What are the implications if it doesn’t work well?
  • What are the implications if it does work well?
  • What are the implications if it works well for some people, but not others?
  • Who benefits and who is harmed?
  • (About datasets) How was it collected? Is it representative?
  • (For any technology) Who is it designed for?

7 of 45

More questions

  • Does the application align with your values?
  • Does the task specification / evaluation metric reflect the things you care about?
  • For recognition tasks:
    • Does the collected training / test set match your true distribution?
  • Are the algorithm’s errors biased?
  • Are you being honest in public descriptions of your results?
  • Is the accuracy/correctness sufficient for public release?

Slide credit: Bharath Hariharan

8 of 45

Bias in computer vision and beyond

  • What follows are a number of examples of bias from the last 100 years

9 of 45

Shirley cards

How Kodak's Shirley Cards Set Photography's Skin-Tone Standard

https://www.npr.org/2014/11/13/363517842/for-decades-kodak-s-shirley-cards-set-photography-s-skin-tone-standard

The Racial Bias Built Into Photography

https://www.nytimes.com/2019/04/25/lens/sarah-lewis-racial-bias-photography.html

Kodak’s Multiracial Shirley Card, North America. 1995.

Example Kodak Shirley Card, 1950s and beyond

10 of 45

Face recognition

  • Probably the most controversial vision technology
  • Three different versions:
    • Face verification: “Is this person Noah Snavely?” (e.g., Apple’s Face Unlock)
    • Face clustering: “Who are all the people in this photo collection”? (e.g., Google Photos search)
    • Face recognition: “Who is this person”? (e.g., identify a person from surveillance footage of a crime scene)
  • Applications can suffer from bias (working well for some populations but not others) and misuse

11 of 45

Google Photos automatic face clustering and recognition

12 of 45

NYT, 12/20/19

https://www.nytimes.com/2019/12/19/technology/facial-recognition-bias.html

13 of 45

New York Times, June 24, 2020

https://www.nytimes.com/2020/06/24/technology/facial-recognition-arrest.html

14 of 45

https://www.nytimes.com/2020/01/18/technology/clearview-privacy-facial-recognition.html

15 of 45

Face analysis

  • Gender classification
  • Age regression
  • Expression classification
  • Ethnicity classification

16 of 45

Gender Shades – Evaluation of bias in Gender Classification

Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. Conference on Fairness, Accountability and Transparency. 2018.

Images from the Pilot Parliaments Benchmark

17 of 45

Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. Conference on Fairness, Accountability and Transparency. 2018.

18 of 45

Case study – upsampling faces

https://github.com/tg-bomze/Face-Depixelizer

PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models

Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin

https://arxiv.org/abs/2003.03808

19 of 45

Case study – upsampling faces

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

20 of 45

Case study – upsampling faces

“We have noticed a lot of concern that PULSE will be used to identify individuals whose faces have been blurred out. We want to emphasize that this is impossible - PULSE makes imaginary faces of people who do not exist, which should not be confused for real people. It will not help identify or reconstruct the original image.

​

We also want to address concerns of bias in PULSE. We have now included a new section in the paper and an accompanying model card directly addressing this bias.”

https://github.com/tg-bomze/Face-Depixelizer, accessed May 4, 2021

21 of 45

Case study – classifying sexual orientation

22 of 45

“We show that faces contain much more information about sexual orientation than can be perceived and interpreted by the human brain… Given a single facial image, a classifier could correctly distinguish between gay and heterosexual men in 81% of cases, and in 74% of cases for women. … Consistent with the prenatal hormone theory of sexual orientation, gay men and women tended to have gender-atypical facial morphology, expression, and grooming styles … our findings expose a threat to the privacy and safety of gay men and women.”

Wang & Kosinski 2017

23 of 45

More questions

  • Does the application align with your values?
  • Does the task specification / evaluation metric reflect the things you care about?
  • For recognition:
    • Does the collected training / test set match your true distribution?
  • Are the algorithm’s errors biased?
  • Are you being honest in public descriptions of your results?
  • Is the accuracy/correctness sufficient for public release?

Slide credit: Bharath Hariharan

24 of 45

Answers

  • Training / test set?
    • 35,326 images from public profiles on a US dating website
  • ”average” images of straight/gay people:
  • Question:
    • Are differences caused by actual differences in faces…
    • … or how people choose to present themselves in dating websites?

​

Slide credit: Bharath Hariharan

25 of 45

Answers

  • Goal: raise privacy concerns.
  • Side-effects?
    • Reinforces potentially harmful stereotypes
    • Provides ostensibly “objective” criteria for discrimination

Slide credit: Bharath Hariharan

26 of 45

Do algorithms reveal sexual orientation or just expose our stereotypes?

Blaise Agüera y Arcas, Alexander Todorov and Margaret Mitchell

https://medium.com/@blaisea/do-algorithms-reveal-sexual-orientation-or-just-expose-our-stereotypes-d998fafdf477

27 of 45

Datasets – Potential Issues

  • Licensing and ownership of data
  • Consent of photographer and people being photographed
  • Offensive content
  • Bias and underrepresentation
    • Including amplifying bias
  • Unintended downstream uses of data

28 of 45

Case study – ImageNet

  • Serious issues with the People subcategory
    • Offensive content, non-visual categories
  • Pointed out by https://excavating.ai/ (Crawford & Paglen, 2019)

29 of 45

30 of 45

Case study – Microsoft Celeb

  • Microsoft Celeb (MS-Celeb-1M): dataset of 10 million face images harvested from the Internet for the purpose of developing face recognition technologies.

​

  • From http://exposing.ai: “While the majority of people in this dataset are American and British actors, the exploitative use of the term ‘celebrity’ extends far beyond Hollywood. Many of the names in the MS Celeb face recognition dataset are merely people who must maintain an online presence for their professional lives: journalists, artists, musicians, activists, policy makers, writers, and academics. Many people in the target list are even vocal critics of the very technology Microsoft is using their name and biometric information to build.”

31 of 45

Case study – Microsoft Celeb

  • Microsoft Celeb taken down May 2019
  • However, dataset still can be found online
  • Case brings up questions of consent and privacy of individuals in a dataset, as well as uses of large-scale face recognition and “runaway datasets”

32 of 45

Case study – LAION-5B

33 of 45

https://techcrunch.com/2024/08/30/the-org-behind-the-data-set-used-to-train-stable-diffusion-claims-it-has-removed-csam/

34 of 45

Some “sunsetted” datasets

  • Microsoft Celeb (MS-Celeb-1M)
  • ImageNet (partial – people category)
  • MIT Tiny Images
  • MegaFace
  • Duke MTMC Dataset

​

  • See https://exposing.ai/datasets/ for more information
  • Additional reference:
    • Large image datasets: A pyrrhic win for computer vision? Vinay Uday Prabhu & Abeba Birhane. https://arxiv.org/abs/2006.16923

35 of 45

Datasheets for Datasets

“The ML community currently has no standardized process for documenting datasets, which can lead to severe consequences in high-stakes domains. To address this gap, we propose datasheets for datasets. In the electronics industry, every component, no matter how simple or complex, is accompanied with a datasheet that describes its operating characteristics, test results, recommended uses, and other information. By analogy, we propose that every dataset be accompanied with a datasheet that documents its motivation, composition, collection process, recommended uses, and so on.”

Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, Kate Crawford. 2018

36 of 45

DeepFakes

https://www.theverge.com/2021/3/5/22314980/tom-cruise-deepfake-tiktok-videos-ai-impersonator-chris-ume-miles-fisher

37 of 45

DeepFakes

https://www.youtube.com/watch?v=nwOywe7xLhs

38 of 45

DeepFakes

  • Active research on both better and better image/video generation and detection of fake images
  • Representative work:

https://peterwang512.github.io/CNNDetection/

39 of 45

Text-to-image models

  • Often trained on datasets that contain copyrighted material

40 of 45

https://www.theverge.com/2023/1/17/23558516/ai-art-copyright-stable-diffusion-getty-images-lawsuit

41 of 45

https://www.technologyreview.com/2022/09/16/1059598/this-artist-is-dominating-ai-generated-art-and-hes-not-happy-about-it/

Example Text-to-image prompt: “Wizard with sword and a glowing orb of magic fire fights a fierce dragon Greg Rutkowski,”

“Dragon Cave”

GREG RUTKOWSKI

42 of 45

https://www.vox.com/future-perfect/23023538/ai-dalle-2-openai-bias-gpt-3-incentives

Generated images of lawyers

Generated images of flight attendants

43 of 45

https://www.nytimes.com/2024/02/22/technology/google-gemini-german-uniforms.html

New York Times, February 22, 2024

44 of 45

Some tools

  • Policy and regulation
    • e.g., a number of cities have banned the use of face recognition by law enforcement
  • Transparency
    • e.g., studies on bias in face recognition have led to reforms by tech companies themselves
    • e.g., datasheets can help downstream users of datasets
  • Awareness (when you conceive of or build a technology, be aware of the questions we’ve discussed)

45 of 45

Questions?