Ken St. Germain

Ken St. Germain, Josh Angichiodo

Correctly identifying the type of file under examination is a critical part of a forensic investigation. The file type alone suggests the embedded content, such as a picture, video, manuscript, spreadsheet, etc. In cases where a system owner might desire to keep their files inaccessible or file type concealed, we propose using an adversarially-trained machine learning neural network to determine a file's true type even if the extension or file header is obfuscated to complicate its discovery. Our semi-supervised generative adversarial network (SGAN) achieved 97.6% accuracy in classifying files across 11 different types. We also compared our network against a traditional standalone neural network and three other machine learning algorithms. The adversarially-trained network proved to be the most precise file classifier especially in scenarios with few supervised samples available. Our implementation of a file classifier using an SGAN is implemented on GitHub (https://ksaintg.github.io/SGAN-File-Classier).

Ken St. Germain, Frank Kragh

Strong authentication in an interconnected wireless environment continues to be an important, but sometimes elusive goal. Research in physical-layer authentication using channel features holds promise as a technique to improve network security for a variety of devices. We propose the use of machine learning and measured multiple-input multiple-output communications channel information to make a decision on whether or not to authenticate a particular device. This work analyzes the use of received channel state information from the wireless environment and demonstrates the employment of a generative adversarial neural network (GAN) trained with received channel data to authenticate a transmitting device. We compared a variety of machine learning techniques and found that the local outlier factor (LOF) algorithm reached 100% accuracy at lower signal to noise ratios (SNR) than other algorithms. However, before LOF reached 100%, we also show that the GAN was more accurate at lower SNR levels.