We present a framework to statistically audit the privacy guarantee conferred by a differentially private machine learner in practice. While previous works have taken steps toward evaluating privacy loss through poisoning attacks or membership inference, they have been tailored to specific models or have demonstrated low statistical power. Our work develops a general methodology to empirically evaluate the privacy of differentially private machine learning implementations, combining improved privacy search and verification methods with a toolkit of influence-based poisoning attacks. We demonstrate significantly improved auditing power over previous approaches on a variety of models including logistic regression, Naive Bayes, and random forest. Our method can be used to detect privacy violations due to implementation errors or misuse. When violations are not present, it can aid in understanding the amount of information that can be leaked from a given dataset, algorithm, and privacy specification.

Because of rampant security breaches in IoT devices, searching vulnerabilities in massive IoT ecosystems is more crucial than ever. Recent studies have demonstrated that control-flow graph (CFG) based bug search techniques can be effective and accurate in IoT devices across different architectures. However, these CFG-based bug search approaches are far from being scalable to handle an enormous amount of IoT devices in the wild, due to their expensive graph matching overhead. Inspired by rich experience in image and video search, we propose a new bug search scheme which addresses the scalability challenge in existing cross-platform bug search techniques and further improves search accuracy. Unlike existing techniques that directly conduct searches based upon raw features (CFGs) from the binary code, we convert the CFGs into high-level numeric feature vectors. Compared with the CFG feature, high-level numeric feature vectors are more robust to code variation across different architectures, and can easily achieve realtime search by using state-of-the-art hashing techniques. We have implemented a bug search engine, Genius, and compared it with state-of-art bug search approaches. Experimental results show that Genius outperforms baseline approaches for various query loads in terms of speed and accuracy. We also evaluated Genius on a real-world dataset of 33,045 devices which was collected from public sources and our system. The experiment showed that Genius can finish a search within 1 second on average when performed over 8,126 firmware images of 420,558,702 functions. By only looking at the top 50 candidates in the search result, we found 38 potentially vulnerable firmware images across 5 vendors, and confirmed 23 of them by our manual analysis. We also found that it took only 0.1 seconds on average to finish searching for all 154 vulnerabilities in two latest commercial firmware images from D-LINK. 103 of them are potentially vulnerable in these images, and 16 of them were confirmed.