AI vs. Our Eye: New Perspectives on Adversial Images | by Martha Larson
- Imagine Interpretation
- Digging deeper into image interpretation
- Is AI for image interpretation "superhuman"?
- Is AI for image interpretation "broken" (due to adversial images)?
1. Image interpretation
State of the art: produce scores over labels for a given image.
Section over convolutional neural networks.
Convolution slides over an image to produce a feature map. The NN learns to use properties of the filters.
- Label classification is of course limited to the labels fed to the network during training; and these labels need to be representative
- And if the label is there, there needs to be enough representative training data available.
Humans do not classify individual objects or describe them first before we can recognize similarity in some scene.
Types of similarity
- Near duplicate similarity
- The actual picture is the same for practical purposes
- This is easy to automate because pixel values will be near identical
- Semantic similarity
- Humans can discuss similarity of similar objects; e.g. two different pictures of pyramids
- This is hard for AI, because the actual pixel values of the image may be very different.
2. Diving deeper
AI can be used on different types of images. Martha distinguishes two groups
- Created Content: created by a person for consumption by other people; to communicate a message
- Online sharing
- News images
- Advertisement images
- Art
- Sensed data: Created by systems, not necessarily for people; also not generally thought to convey a message
- Surveillance footage
- Biometric authentication
- Medical Imaging
- Camera feed of robots or vehicles
An important difference here is that for created content we do not have direct access to the interpretations of that content by creators and users. In the second case we have a lot of control about how the images are created and how they can be used.
This difference is important for AI because it influences how we
- Collect training data
- Captured: Since you have access to the system, it's easy to capture more data as needed. E.g. let your automatic vehicle drive another few laps
- Created: you are reliant on what people produce naturally; and e.g. advertisements and such really produce novel content; novelty is interesting anyways, because for reliable training data you want many pictures of the same thing from different perspectives
- Label training data
- Captured: Because the system is controlled, you often know the ground truth, or experts can fill it in.
- Created: Labels are by definition incomplete, because an open-ended system needs to serve most people, but you can't ask all of them to provide labels.
- Created: also there's simply too much to label. E.g. 300 mil pictures per day uploaded to Facebook. Theoretically, we would need similarity judgments between all pairs.
- Evaluate the AI system
- Captured: Success or failure is well-defined, e.g. a clinical test will tell if that lung scan indeed showed cancer
- Created: Because trained-labels are incomplete, long-term consequences of the system cannot be completely foreseen
An AI System Developer is not an expert
We usually do not claim to be experts when reading lung X-rays (unless we have extensive medical training). But it's very tempting to think we are experts on the interpretation of an average image, while in fact this requires context, knowledge, relevant life experiences, or e.g. whether we took the image. So it's not simply the case that an AI expert can fully reliably label images created by people. Multiple perspectives are needed in data selection and labeling. Complete coverage is not possible. And control should remain in the hands of the content creators and system users.
Is AI for image interpretation "superhuman"?
We can said AI image recognition is superhuman in certain ways:
- It can classify details in images in ways that we cannot or find very hard
- It can classify several things we also classify, but do it tirelessly at a large scale and high speeds
Martha argues we often think about the first point, but overlook the second; whereas the second may have a lot of consequences.
E.g. because data can be processed at large scale, one may for example immediately geo-tag a particular location because there have previously been many pictures made by others at the same location.
But this poses a problem, because this easily leads to privacy violations.
Adversial images: is AI broken?
We are not talking about adversial learning here, but adversial examples.
How do we generate adversial examples?
We mimimize e.g. the Euclidean distance between the adversial image and the original, such that the original is pushing over the original decision boundary. I.e. you want it to be misclassified, while at the same time still looking like the original.
We can also have adversial examples w.r.t. similarity measures or retrieval through image queries. Here you are not taking into account a target class, you just care about pushing it away from its original position in the ranking.
Adversial techniques will hinder, if not prevent, malicious use of our images. E.g. it could help us protect from exposing our location in automatic geo-tagging based on the background in a picture. This method also still works if pics are breached or leaked.
Adversial images can also be dangerous for typical sensed-data AI systems, such as self-driving cars. These systems have a purpose; so adversial images break this purpose. (May be good or bad; but this "breaking" is at least well-defined here).
Adversial images for created-content AI systems can be used to protect users and are not necessarily bad. Images that are adversial w.r.t. human viewers instead of AI-systems can be seen as a natural part of communication (imitation is natural and often funny); deep fakes are problematic though, but even they wouldn't be there there would still be shallow fakes and cheap fakes.