For model privacy, local model parameters in federated learning shall be obfuscated before sent to the remote aggregator. This technique is referred to assecure aggregation. However, secure aggregation makes model poisoning attacks such as backdooring more convenient given that existing anomaly detection methods mostly require access to plaintext local models. This paper proposes a new federated learning technique SAFE-Learning to support backdoor detection for secure aggregation. We achieve this through two new primitives -oblivious random grouping (ORG)andpartial parameter disclosure (PPD). ORG partitions participants into one-time random subgroups with group configurations oblivious to participants; PPD allows secure partial disclosure of aggregated subgroup models for anomaly detection without leaking individual model privacy. ORG is based on our construction of several new primitives including tree-based random subgroup generation, oblivious secure aggregation, and randomized Diffie-Hellman key exchange. ORG can thwart colluding attackers from knowing each other’s group membership assignment with non-negligible advantage than random guess. Backdoor attacks are detected based on statistical distributions of the subgroup aggregated parameters of the learning iterations. SAFELearning can significantly reduce backdoor model accuracy without jeopardizing the main task accuracy under common backdoor strategies. Extensive experiments show SAFELearning is robust against malicious and faulty participants, whilst being more efficient than the state-of-art secure aggregation protocol in terms of both communication and computation costs.

Real-time decision making with unmanned aerial vehicles (UAVs) imagery is desired in many applications. Deep learning (DL) is a promising enabler for such applications thanks to its recent advancements. However, direct execution of DL models on UAVs, especially small and micro ones, would not only introduce severe delay but also significantly shorten the flight time of UAVs due to the high energy consumption. Realtime transmission of UAV images to ground edge devices for deep analysis can mitigate the computational complexity but may introduce severe interference to ground devices, in addition unpredictable delays due to the dynamic network conditions. To minimize real-time image transmission, this paper designs a new offloading prediction algorithm which first estimates nearfuture need for DL of each UAV and transmit images only when necessary. Holistic resource allocation is made at the edge based on the offloading likelihood analysis of multiple UAVs as well as available resources. Experimental results on real UAV video clips show that our design can save 92% of the communication costs with less than 4% false positive rate.

Due to the inherent broadcast nature of the wireless medium, Wireless Local Area Networks (WLANs) are targets of a variety of malicious attacks, for example, MAC identity spoofing, rogue AP attack, and network freeloading. These attacks invite security and privacy threats and hinder the worry-free deployment of WLAN networks. To thwart these attacks, existing research has proposed to use hardware-specific imperfections as a unique unforgeable fingerprint for the APs and/or clients. Unfortunately, existing solutions are limited to static and stable environments or use customized hardware preventing their wide-scale adoption. To overcome the limitations, in this work, we propose to use the distribution of relative phase differences between MIMO-radio transmitter oscillators as a distinguishing trait or fingerprint. More specifically, we show that the nonidealities of the multiple RF chains on a single MIMO-OFDM (Multiple Input Multiple Output-Orthogonal Frequency Division Multiplexing) transmitter can be extracted and utilized as a reliable device fingerprint. Each transmitter RF chain has a random initial phase offset, and their difference relative to one another is stable over time, differs uniquely for each transmitter device and cannot be altered by the adversary without significant effort and cost. Our functional prototype measures these unknown phase differences using PHY-layer Channel State Information (CSI) of the in-band channel obtained from off-the-shelf hardware. Our design eliminates expensive custom-built hardware, is invariant to environmental variations and supports device mobility making it practical and deployable in real indoor settings. Experimental evaluation using 17 Intel Network Interface Cards (NICs) resulted in 97 % and 92 % device identification accuracy for static and mobile device states respectively. Such promising results with identical model and manufacturer devices wherein underlying manufacturing variations are typically low showcase the effectiveness of our design and suggest even higher accuracy across multi-model and multi-manufacturer cards because of the higher manufacturing variations.