Black-Box Adversarial Attacks via Markov Random Field Covariance

The patent performs preliminary action by pre-estimating the gradient covariance matrix using a small subset of training data before the actual adversarial attack. This pre-computed covariance information is then reused during the attack phase, eliminating the need for repeated gradient estimations in each attack step. The covariance estimation is done once beforehand, and the resulting matrix guides the entire attack process, thereby reducing both query efficiency loss and time consumption.

The patent applies beforehand cushioning by pre-computing and storing the gradient covariance matrix as a form of preparatory information that cushions against the need for repeated expensive operations. This pre-computed statistical information acts as a buffer that enables efficient attack generation without requiring real-time gradient computations, thus cushioning the system against query inefficiency and time loss.

The patent applies parameter changes by transforming the gradient estimation problem into a covariance matrix estimation problem. Instead of estimating full gradient vectors which are high-dimensional and complex, the method estimates the covariance matrix parameters that capture gradient correlations. This parameter transformation reduces the computational burden while preserving the essential correlation information, thereby improving attack accuracy without proportionally increasing computation complexity.

The patent applies local quality by focusing computational resources on estimating only the relevant covariance structure rather than full gradient information. The method selectively captures the correlation patterns that are most important for adversarial attack generation, rather than uniformly processing all gradient dimensions. This localized approach to quality estimation improves accuracy where it matters most while controlling overall computational complexity.