AI Sec Watch

Adversaries can hide a smaller image within a larger one so that it becomes visible when a computer resizes the image using insecure interpolation (a method of calculating pixel values between known points). The video demonstrates this attack technique and explains how to prevent it from happening.

ParlAI, a framework for training AI models on dialogue datasets, has a vulnerability where it unsafely loads YAML files (a data format), allowing attackers to execute arbitrary code on affected systems. The vulnerability occurs because the framework uses an unsafe YAML loader that can be tricked into running malicious code hidden in data files.

This post describes Microsoft Counterfit, a tool for testing machine learning models against adversarial attacks (subtle modifications to input data designed to fool AI systems). The author demonstrates how to set up Counterfit, create a custom target for a husky image classifier, and use the tool's built-in attack modules to test the model's robustness.

TensorFlow, an open-source machine learning platform, had a bug where certain shape functions created temporary data structures (ShapeAndType structs) that were deleted too quickly, causing crashes (segfaults, or sudden program failures) if other code tried to access them. The issue was that while normal output shapes were being protected by copying them to safer ownership, the code wasn't doing the same protection for shapes and types together.

TensorFlow (an open source machine learning platform) had a bug where Go code could crash the program during memory cleanup of string tensors if encoding failed. The problem occurred because the cleanup process assumed encoding always succeeded, but didn't check whether it actually did.

TensorFlow, an open-source machine learning platform, has a vulnerability where an attacker can create a specially crafted TFLite model (a lightweight version of TensorFlow for mobile and embedded devices) that causes a division by zero error (a crash that happens when code tries to divide a number by zero) in its LSH projection feature. This flaw affects multiple versions of TensorFlow.

TensorFlow Lite (TFLite, a lightweight version of TensorFlow for mobile and embedded devices) has a vulnerability in its `GatherNd` and `Gather` operations that fail to check for negative indices. An attacker can exploit this by creating a specially designed model with negative values to read sensitive data from the heap (temporary memory storage), potentially exposing private information.

TensorFlow, an open source machine learning platform, has a vulnerability in TFLite (TensorFlow Lite, a lightweight version for mobile devices) where a negative `axis` parameter value can cause the software to read data outside the intended memory area. This could potentially expose sensitive information or crash the program.

TensorFlow (an open source platform for machine learning) has a vulnerability in its pooling operations where the code doesn't check if divisors are zero before dividing, which can cause crashes. The issue has been patched and will be included in upcoming versions of TensorFlow.