22-W12 Research Seminar

AI-synthesized faces are indistinguishable from realfaces and more trustworthy

[Paper]

Abstract

Artificial intelligence (AI)–synthesized text, audio, image, and video are being weaponized for the purposes of nonconsensual intimate imagery, financial fraud, and disinformation campaigns. Our evaluation of the photorealism of AI-synthesized faces indicates that synthesis engines have passed through the uncanny valley and are capable of creating faces that are indistinguishable—and more trustworthy—than real faces.

 

모델에 대한 논문이라기 보다는 Image Generation에 대한 Deep Learning 모델의 결과물이 Uncanny vally를 넘어섰다는 것을 보여주기 위한 Research Paper에 가깝다.

 

that synthesis engines have passed through the uncanny valley and are capable of creating faces that are indistinguishable

 

 

uncanny vally

 

 

 

실제로 GAN등의 Image Synethesize 모델을 사용해서 생성한 얼굴과 실제 사람의 사진을 비교했을 때, 단순히 '비슷하다'가 아닌 '구분하기 힘들다'(indistinguishable)로 결론이 나고 있으며, 이로 인한 문제들(associated risks)가 이점보다 클 수 있음을 고려해야 한다고 결론내린다. 

 

Synthetically generated faces are not just highly photorealistic, they are nearly indistinguishable from real faces and are judged more trustworthy

We, therefore, encourage those developing these technologies to consider whether the associated risks are greater than their benefits.

we encourage the graphics and vision community to develop guidelines for the creation and distribution of synthetic media technologies that incorporate ethical guidelines for researchers, publishers, and media distributors.

 

The four most ( Top ) and four least ( Bottom ) trustworthy faces and their trustworthy rating on a scale of 1 (very untrustworthy) to 7 (very trustworthy). Synthetic faces (S) are, on average, more trustworthy than real faces (R).

 

 

FYI. 실제로 이 Research Paper에서는 SOTA StyleGAN2를 사용해서 이미지를 생성했다고 한다. 

We selected 400 faces synthesized using the state-of-the-art StyleGAN2, ensuring diversity across gender (200 women; 200 men), estimated age (ensuring a range of ages from children to older adults), and race (100 African American or Black, 100 Caucasian, 100 East Asian, and 100 South Asian).

 

Can Neural Nets Learn the Same Model Twice? Investigating Reproducibility and Double Descent from the Decision Boundary Perspective

[Paper]

Abstract

We discuss methods for visualizing neural network decision boundaries and decision regions. We use these visualizations to investigate issues related to reproducibility and generalization in neural network training. We observe that changes in model architecture (and its associate inductive bias) cause visible changes in decision boundaries, while multiple runs with the same architecture yield results with strong similarities, especially in the case of wide architectures.

We also use decision boundary methods to visualize double descent phenomena. We see that decision boundary reproducibility depends strongly on model width. Near the threshold of interpolation, neural network decision boundaries become fragmented into many small decision regions, and these regions are non-reproducible. Meanwhile, very narrows and very wide networks have high levels of reproducibility in their decision boundaries with relatively few decision regions. We discuss how our observations relate to the theory of double descent phenomena in convex models.

 

Do different neural architectures have measurable differences

changes in model architecture (and its associate inductive bias) cause visible changes in decision boundaries, while multiple runs with the same architecture yield results with strong similarities, especially in the case of wide architectures.

 

• Convolutional model들의 경우 비슷한 decision boundaries를 갖는 것을 확인할 수 있다.
• 반면 FC, ViT, MLP Mixer등은 각 모델간의 decision boundaries가 크게 다르다. 
• Convolutional model들의 경우 high reproducibilities를 가진다. (wide architectures), 즉 trial을 반복해도 얻게되는 decision boundaries에는 큰 차이가 없다.

 

 

 

Double Descents

We show that the double descent phenomenon occurs in CNNs, ResNets, and transformers: performance first improves, then gets worse, and then improves again with increasing model size, data size, or training time. This effect is often avoided through careful regularization.

 

CNN, ResNet등의 모델의 Training 과정에서 나타나는 현상으로 퍼포먼스가 초기에는 증가하다가 모델 사이즈가 증가하면서(여기서 모델 사이즈란, 레이어의 갯수를 의미함) 퍼포먼스가 급격하게 떨어지고, 그러다가 어느 임계점을 지나면서 퍼포먼스가 급격하게 좋아지는 현상을 의미한다. 

 

 

 

 

Noisy labels increases instability

실제로 학습을 시켜보면 Noisy Label이 포함된 경우 width Parameter K가 interpolation threshold에 가까워질수록 decision boundary가 점점 더 fragmented되는 현상을 확인할 수 있으며, 해당 threshold을 넘어서게 되면, 실제로 성능이 좋아지는 것을 확인할 수 있다. 즉 모델 학습에 있어서 noisy label이 들어가게 되면 Double Descent 현상을 일으키게 되며, 모델의 불안정성(instability)을 높인다는 것이다.

 

 

이에 대해 저자들이 강조하는 포인트를 크게 2가지 관점에서 생각해볼 수 있고 두 가지 관점이 연관되었다고 볼 수 있으나 두번째 관점에 조금 더 힘을 싣고 있는 듯 하다.

1. mislabled된 데이터(noisy label)를 학습했으므로, data space에서 mislabeled point에 해당하는 부분을 다른 색으로 칠해야한다. 즉, 주어진 데이터를 가지고 학습은 잘 됐는데 주어진 데이터가 문제였다
2. 애초에 학습 자체가 잘 되지 않았다. mislabeled point는 기존의 패턴에 벗어난 것이고, 이로 인해서 네트워크 자체가 제대로 학습되지 않았다.

 

 

 

 

Reference

https://creamnuts.github.io/short%20review/deep_double_descent/

Deep Double Descent: Where Bigger Models and More Data Hurt

https://openai.com/blog/deep-double-descent/?fbclid=IwAR0QgePqoVIfDOlCDDZYDX50sbR-NTmy4c_1sbe79jPCfw57L1CMcquZkaY 

Deep Double Descent