Implementations of SGD on distributed systems create new vulnerabilities, which can be identified and misused by one or more adversarial agents. Recently, it has been shown that well-known Byzantine-resilient gradient aggregation schemes are indeed vulnerable to informed attackers that can tailor the attacks (Fang et al., 2020; Xie et al., 2020b). We introduce MixTailor, a scheme based on randomization of the aggregation strategies that makes it impossible for the attacker to be fully informed. Deterministic schemes can be integrated into MixTailor on the fly without introducing any additional hyperparameters. Randomization decreases the capability of a powerful adversary to tailor its attacks, while the resulting randomized aggregation scheme is still competitive in terms of performance. For both iid and non-iid settings, we establish almost sure convergence guarantees that are both stronger and more general than those available in the literature. Our empirical studies across various datasets, attacks, and settings, validate our hypothesis and show that MixTailor successfully defends when well-known Byzantine-tolerant schemes fail.
翻译:在分布式系统上实施 SGD 实施分布式系统中的 SGD 造成新的脆弱性,可由一个或多个对抗性代理人识别和滥用。最近,已经表明,众所周知的Byzantine抗逆梯度梯度汇总计划确实容易受到知情袭击者的影响,而袭击者能够根据具体情况调整袭击(Fang等人,2020年;Xie等人,2020年b),我们引入了MixTailor计划,这是一个基于集成战略随机化的计划,使袭击者无法充分了解袭击者的情况。确定性计划可以纳入飞行上的MixTailor,而不引入任何额外的超参数。随机化降低了强大的对手调整其袭击的能力,而由此产生的随机集成计划在性能方面仍然具有竞争力。对于igid和非iid环境,我们几乎肯定了趋同性保证,这些保证比文献中已有的更加有力和普遍。我们在各种数据集、攻击和环境中进行的经验研究,证实了我们的假设,并表明MixTailor在众所周知的Byzant耐受袭击计划失败时成功地进行辩护。