Biblio

Modern detection systems use sensor outputs available in the deployment environment to probabilistically identify attacks. These systems are trained on past or synthetic feature vectors to create a model of anomalous or normal behavior. Thereafter, run-time collected sensor outputs are compared to the model to identify attacks (or the lack of attack). While this approach to detection has been proven to be effective in many environments, it is limited to training on only features that can be reliably collected at detection time. Hence, they fail to leverage the often vast amount of ancillary information available from past forensic analysis and post-mortem data. In short, detection systems do not train (and thus do not learn from) features that are unavailable or too costly to collect at run-time. Recent work proposed an alternate model construction approach that integrates forensic "privilege" information–-features reliably available at training time, but not at run-time–-to improve accuracy and resilience of detection systems. In this paper, we further evaluate two of proposed techniques to model training with privileged information: knowledge transfer, and model influence. We explore the cultivation of privileged features, the efficiency of those processes and their influence on the detection accuracy. We observe that the improved integration of privileged features makes the resulting detection models more accurate. Our evaluation shows that use of privileged information leads to up to 8.2% relative decrease in detection error for fast-flux bot detection over a system with no privileged information, and 5.5% for malware classification.

Machine learning (ML) models, e.g., deep neural networks (DNNs), are vulnerable to adversarial examples: malicious inputs modified to yield erroneous model outputs, while appearing unmodified to human observers. Potential attacks include having malicious content like malware identified as legitimate or controlling vehicle behavior. Yet, all existing adversarial example attacks require knowledge of either the model internals or its training data. We introduce the first practical demonstration of an attacker controlling a remotely hosted DNN with no such knowledge. Indeed, the only capability of our black-box adversary is to observe labels given by the DNN to chosen inputs. Our attack strategy consists in training a local model to substitute for the target DNN, using inputs synthetically generated by an adversary and labeled by the target DNN. We use the local substitute to craft adversarial examples, and find that they are misclassified by the targeted DNN. To perform a real-world and properly-blinded evaluation, we attack a DNN hosted by MetaMind, an online deep learning API. We find that their DNN misclassifies 84.24% of the adversarial examples crafted with our substitute. We demonstrate the general applicability of our strategy to many ML techniques by conducting the same attack against models hosted by Amazon and Google, using logistic regression substitutes. They yield adversarial examples misclassified by Amazon and Google at rates of 96.19% and 88.94%. We also find that this black-box attack strategy is capable of evading defense strategies previously found to make adversarial example crafting harder.