Roundup Welcome to your weekly dose of AI-related news beyond what we’ve already covered.
Kubernetes machine-learning toolkit exploited to mine crypto: Microsoft said this week it discovered a number of Kubernetes clusters infected with cryptomining code, which was injected into these systems via the Kubeflow machine-learning toolkit.
It appears administrators of the clusters poorly secured their installations, exposing their Kubeflow dashboards to the public internet. This front-door access was abused by miscreants to deploy an image called ddsfdfsaadfs/dfsdf: 99 from a public repository as a container in the cluster, which quietly set up and executed the digital-cash miner.
Kubeflow is a framework that runs TensorFlow models on Kubernetes. It’s a good target for miscreants hoping to mine cryptocurrency as it runs in clusters that often have a lot of compute resources, including GPU acceleration.
“[Kubeflow] is a containerized service,” Microsoft Azure Security Center’s Yossi Weizman explained in the above-linked advisory. “Therefore, if attackers somehow get access to Kubeflow, they have multiple ways to run their malicious image in the cluster.
“The attacker used an exposed dashboard (Kubeflow dashboard in this case) for gaining initial access to the cluster. The execution and persistence in the cluster were performed by a container that was deployed in the cluster. The attacker managed to move laterally and deploy the container using the mounted service account. Finally, the attacker impacted the cluster by running a cryptocurrency miner.”
The takeaway message here is: secure your control panels with authentication and access policies, before someone hijacks them to mine coins on your own dime. And monitor the containers deployed in your clusters.
Is Clearview AI’s facial-recognition being used during America’s civil unrest? That’s what US Senator Edward Markey (D-MA) wants to know. Markey wrote a letter to Clearview’s CEO Hoan Ton-That asking if his upstart’s facial-recognition software has been used by law enforcement while countless folks across America protest against police brutality and systemic racism.
“In light of the ongoing protests and demonstrations across the country, I write with additional questions and to reiterate the need for your company to take urgent action to prevent the harmful use of its product,” the letter read.
Using facial recognition to scan crowds, allowing the cops and Feds to identify who was present during the demonstrations, could deter folks from going out to protest, and infringes upon their First Amendment rights, Markey argued. Clearview trained its AI product on billions of images scraped from social media pages, and sells it to law enforcement so that officers and agents can trace still images of people’s faces to their online profiles and identities.
“Will you commit to explicitly prohibiting law enforcement agencies or others from using Clearview AI’s technology to monitor or identify peaceful protestors? If so, please detail how you will do so. If not, why not?” Markey asked.
It’s not the first time Markey has tried to grill the controversial startup. The senator has sent multiple letters to Ton-That before, and, unsurprisingly, the CEO has never really been that forthcoming in his answers. Clearview has refused to reveal who its clients are, and has not submitted its software to be audited by a third party to check its accuracy and any potential racial biases.
The Register asked Clearview for comment.
Tens and thousands of faces scanned at college football game: Here’s another unsettling facial-recognition use case. A company known as VSBLTY used four cameras to take pictures of thousands of people who attended a college football game between the Oregon Ducks and the Wisconsin Badgers at the Rose Bowl Stadium in California at the start of this year.
VSBLTY straddles “the intersection of marketing and security,” according to its website. It essentially sells cameras and software to sift through crowds of people to identify “persons of interest,” and can analyse footage to determine people’s gender, age, emotional sentiment, and whether they looked at ads on the display boards around a stadium. The cameras at the American football game were thus apparently deployed for advertising reasons, OneZero’s Dave Gershgorn first reported.
“The VSBLTY technology counted individual fans, documented both age and sex, recorded what they watched on video screens and for how long, collected 30,000 impressions, as well as the percentage of views per impression,” the biz said. “The test also showed who watched what commercial messages, for how long, and what kind of information attracted and held interest.”
We note that 90,000 people attended the match. VSBLTY believes that by collecting this type of data, sports venues can better place their adverts, or can prove that some people are paying attention to them at least.
“Traffic count and other venue data collected, when combined with machine learning, can help improve operational efficiencies and venue logistics. Facts about fans, their habits and actions — in addition to demographic and psychographic information — will help plan audience activities as well as serve as a tool to validate the value of on-site advertising impressions to sponsors,” said CEO Jay Hutton.
It’s unknown if VSBLTY got the explicit consent of fans attending the match.
Facebook’s Detectron2 used for the other sort of mining: Facebook this week gushed about its open-source object-recognition model >Detectron2 being used by Australian companies Datarock and Solve Geosolutions to hunt for minerals underground.
Detectron2 was launched last year, and has become one of Facebook’s most popular open-source projects. “Utilizing Detectron2, Datarock’s developers were able to create models capable of conducting rock quality designation (a metric used to understand the strength of rock), rock fracture prediction, and other high-value but difficult-to-collect geological analyses,” Facebook said.
By understanding the local geology, mining companies can find sites that are worth digging into, such as ones that contain higher quality ores. Geologists and engineers need to study a good number of potential sites, though, which can take a long time and is often pretty subjective. Datarock and Solve Geosolutions teamed up to automate part of this process with machine learning.
Detectron2 is written in PyTorch and has been useful for analyzing rock data thanks to its ability to segment images into different parts to depict various types and sizes of rock fragments.
If you’re curious about using Detectron2 for some of your computer vision projects, here’s the code. ®