Algorithmic discrimination is one of the significant concerns in applying machine learning models to a real-world system. Many researchers have focused on developing fair machine learning algorithms without discrimination based on legally protected attributes. However, the existing research has barely explored various security issues that can occur while evaluating model fairness and verifying fair models. In this study, we propose a fairness audit framework that assesses the fairness of ML algorithms while addressing potential security issues such as data privacy, model secrecy, and trustworthiness. To this end, our proposed framework utilizes confidential computing and builds a chain of trust through enclave attestation primitives combined with public scrutiny and state-of-the-art software-based security techniques, enabling fair ML models to be securely certified and clients to verify a certified one. Our micro-benchmarks on various ML models and real-world datasets show the feasibility of the fairness certification implemented with Intel SGX in practice. In addition, we analyze the impact of data poisoning, which is an additional threat during data collection for fairness auditing. Based on the analysis, we illustrate the theoretical curves of fairness gap and minimal group size and the empirical results of fairness certification on poisoned datasets.

Recent research on Internet traffic classification has yield a number of data mining techniques for distinguishing types of traffic, but no systematic analysis on "Why" some algorithms achieve high accuracies. In pursuit of empirically grounded answers to the "Why" question, which is critical in understanding and establishing a scientific ground for traffic classification research, this paper reveals the three sources of the discriminative power in classifying the Internet application traffic: (i) ports, (ii) the sizes of the first one-two (for UDP flows) or four-five (for TCP flows) packets, and (iii) discretization of those features. We find that C4.5 performs the best under any circumstances, as well as the reason why; because the algorithm discretizes input features during classification operations. We also find that the entropy-based Minimum Description Length discretization on ports and packet size features substantially improve the classification accuracy of every machine learning algorithm tested (by as much as 59.8%!) and make all of them achieve >93% accuracy on average without any algorithm-specific tuning processes. Our results indicate that dealing with the ports and packet size features as discrete nominal intervals, not as continuous numbers, is the essential basis for accurate traffic classification (i.e., the features should be discretized first), regardless of classification algorithms to use.

One of the key challenges in designing a rate adaptation scheme for IEEE 802.11 wireless LANs (WLANs) is to differentiate bit errors from link-layer collisions. Many recent rate adaptation schemes adopt the RTS/CTS mechanism to prevent collision losses from triggering unnecessary rate decrease. However, the RTS/CTS handshake incurs significant overhead and is rarely activated in today's infrastructure WLANs. In this paper we propose a new rate adaptation scheme that mitigates the collision effect on the operation of rate adaptation. In contrast to previous approaches adopting fixed rate-increasing and decreasing thresholds, our scheme varies threshold values based on the measured network status. Using the "retry" information in 802.11 MAC headers as feedback, we enable the transmitter to gauge current network state. The proposed rate adaptation scheme does not require additional probing overhead incurred by RTS/CTS exchanges and can be easily deployed without changes in firmware. We demonstrate the effectiveness of our solution by comparing with existing approaches through extensive simulations.