The sticker “allows attackers to create a physical-world attack without prior knowledge of the lighting conditions, camera angle, type of classifier being attacked, or even the other items within the scene.” So, after such an image is generated, it could be “distributed across the Internet for other attackers to print out and use.”

This is why many AI researchers are worried about how these methods might be used to attack systems like self-driving cars. Imagine a little patch you can stick onto the side of the motorway that makes your sedan think it sees a stop sign, or a sticker that stops you from being identified up by AI surveillance systems. “Even if humans are able to notice these patches, they may not understand the intent [and] instead view it as a form of art,” the researchers write.

Here is a 3D-printed turtle that is classified at every viewpoint as a “rifle” by Google’s InceptionV3 image classifier, whereas the unperturbed turtle is consistently classified as “turtle”.

We do this using a new algorithm for reliably producing adversarial examples that cause targeted misclassification under transformations like blur, rotation, zoom, or translation, and we use it to generate both 2D printouts and 3D models that fool a standard neural network at any angle. Our process works for arbitrary 3D models - not just turtles! We also made a baseball that classifies as an espresso at every angle! The examples still fool the neural network when we put them in front of semantically relevant backgrounds; for example, you’d never see a rifle underwater, or an espresso in a baseball mitt.

in today’s paper Moosavi-Dezfooli et al., show us how to create a _single_ perturbation that causes the vast majority of input images to be misclassified.

[The] results suggest that classifiers based on modern machine learning techniques, even those that obtain excellent performance on the test set, are not learning the true underlying concepts that determine the correct output label. Instead, these algorithms have built a Potemkin village that works well on naturally occuring data, but is exposed as a fake when one visits points in space that do not have high probability in the data distribution.

The NSA evaluates the SKYNET program using a subset of 100,000 randomly selected people (identified by their MSIDN/MSI pairs of their mobile phones), and a a known group of seven terrorists. The NSA then trained the learning algorithm by feeding it six of the terrorists and tasking SKYNET to find the seventh. This data provides the percentages for false positives in the slide above.

"First, there are very few 'known terrorists' to use to train and test the model," Ball said. "If they are using the same records to train the model as they are using to test the model, their assessment of the fit is completely bullshit. The usual practice is to hold some of the data out of the training process so that the test includes records the model has never seen before. Without this step, their classification fit assessment is ridiculously optimistic."

The reason is that the 100,000 citizens were selected at random, while the seven terrorists are from a known cluster. Under the random selection of a tiny subset of less than 0.1 percent of the total population, the density of the social graph of the citizens is massively reduced, while the "terrorist" cluster remains strongly interconnected. Scientifically-sound statistical analysis would have required the NSA to mix the terrorists into the population set before random selection of a subset—but this is not practical due to their tiny number.

This may sound like a mere academic problem, but, Ball said, is in fact highly damaging to the quality of the results, and thus ultimately to the accuracy of the classification and assassination of people as "terrorists." A quality evaluation is especially important in this case, as the random forest method is known to overfit its training sets, producing results that are overly optimistic. The NSA's analysis thus does not provide a good indicator of the quality of the method.

The problem with this is that once you accept this framing, and note the happy coincidence that your paymasters just happen to have found a way to spy on everyone, the conclusion is obvious: just mine all of the data, from everyone to everyone, and use an algorithm to figure out who’s guilty. The bad guys have a Modus Operandi, as anyone who’s watched a cop show knows. Find the MO, turn it into a data fingerprint, and you can just sort the firehose’s output into ”terrorist-ish” and ”unterrorist-ish.”

Once you accept this premise, then it’s equally obvious that the whole methodology has to be kept from scrutiny. If you’re depending on three ”tells” as indicators of terrorist planning, the terrorists will figure out how to plan their attacks without doing those three things.

This even has a name: Goodhart's law. "When a measure becomes a target, it ceases to be a good measure." Google started out by gauging a web page’s importance by counting the number of links they could find to it. This worked well before they told people what they were doing. Once getting a page ranked by Google became important, unscrupulous people set up dummy sites (“link-farms”) with lots of links pointing at their pages.

As a library, SAMOA contains state-of-the-art implementations of algorithms for distributed machine learning on streams. The first alpha release allows classification and clustering. For classification, we implemented a Vertical Hoeffding Tree (VHT), a distributed streaming version of decision trees tailored for sparse data (e.g., text). For clustering, we included a distributed algorithm based on CluStream. The library also includes meta-algorithms such as bagging.