AntiPure: Multifaceted Purification Methods

• AntiPure is a multi-interpretational term in security and machine learning that denotes methods for dataset, adversarial, and anti-purification.
• In Android malware detection, AntiPure employs Positive and Unlabeled learning to remove mislabeled contaminants, dramatically boosting true positive rates under heavy contamination.
• In adversarial defenses, AntiPure-inspired systems leverage diffusion models, stochastic sampling, and latent guidance to stress-test and improve purification pipelines.

AntiPure is a non-uniform term in recent security and machine-learning literature. In one lineage, it denotes a PUDROID-inspired contaminant-removal module for Android malware detection that purifies a presumed benign training pool by identifying undisclosed malware with Positive and Unlabeled learning. In another, it denotes a diagnostic protective perturbation designed to remain effective within a diffusion-based purification–customization workflow. In broader adversarial-ML usage, the term also appears as a contextual label for purification-oriented defenses and for analyses that stress-test or invalidate purification pipelines (Sun et al., 2017, Yang et al., 17 Sep 2025, Lin et al., 2024, Kassis et al., 2024).

1. Terminological scope

The supplied literature presents AntiPure in three distinct senses. The first is dataset purification: a preprocessing stage that removes contaminants from a training corpus before a classifier is learned. The second is adversarial purification: a purifier placed in front of a classifier so that perturbed inputs are mapped back toward the clean data manifold. The third is anti-purification: perturbations or evaluation procedures designed to persist through purification or to expose its failure modes (Sun et al., 2017, Lin et al., 2024, Yang et al., 17 Sep 2025, Kassis et al., 2024).

This terminological spread matters because the underlying object of purification differs across domains. In Android malware detection, the target is a mislabeled sample inside a nominally benign dataset. In adversarial image defense, the target is an attack-induced perturbation added at test time. In anti-purification work, the target is the purifier itself: either by constructing inputs that survive denoising or by showing that robustness claims rest on flawed gradients or invalid evaluation protocols. A plausible implication is that “AntiPure” is best understood as a family resemblance term rather than a single standardized method.

2. AntiPure as contaminant removal in Android malware detection

In the Android malware setting, AntiPure is described as a contaminant-removal module inspired by PUDROID, or “Positive and Unlabeled learning-based malware detection for Android,” with the specific goal of purifying the “benign” training set by automatically identifying and removing mislabeled malware. The motivating problem is that trusted stores can distribute undisclosed malware, repackaging is a dominant attack vector, and contaminants flatten class separation for supervised learners. The consequence reported in the supplied data is substantial degradation in true positive rate, accuracy, and F1 under heavy contamination; for example, baseline Random Forest TPR falls to 8.69% at a 3:1 contaminant-to-malware ratio, while the PU-based purifier restores performance (Sun et al., 2017).

The pipeline begins with a positive set $P$ of known malware and an unlabeled set $U$ of presumed benign apps. Static and network features are extracted from permissions, API usage, and URLs converted to IP addresses; activities, services, raw URLs, and high-cardinality intents or components are intentionally excluded. Feature selection then applies dataset-size–aware thresholds governed by

$$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$

with $\eta=2$ in the reported experiments. This reduces the feature space to about 2,200 features, described as 93% fewer than Drebin’s roughly 300,000 features, while preserving class separability (Sun et al., 2017).

Its core learning stage uses Positive and Unlabeled formalism. Samples in $P$ are assigned discovery label $z(s)=1$ and hidden malware label $y(s)=1$, while samples in $U$ have $z(s)=0$ and $y(s)\in\{0,1\}$. The stated key constraint is

$U$0

and the Elkan–Noto-style assumption is

$U$1

A probabilistic base classifier learns $U$2, after which the malware posterior is scaled as

$U$3

with $U$4 estimated by

$U$5

The AntiPure malware scorer is then

$U$6

and samples in $U$7 with $U$8 are flagged as contaminants and removed (Sun et al., 2017).

The reported empirical effect is strongest under heavy contamination. At a 1:1 contaminant-to-malware ratio, Random Forest TPR rises from 44.25% without PU correction to 83.32% with PU correction; at 3:1, it rises from 8.69% to 71.51%; and at 8:1, from 2.14% to 52.14%. In reverse contamination, where benign apps are mixed into the malware set at an 8:1 benign-to-malware ratio, AntiPure with Random Forest reaches 96.36% accuracy, whereas Random Forest without PU achieves 55.5% and Decision Tree without PU remains near random performance. The supplied data also emphasizes that Random Forest is more robust than SVM under extreme contamination (Sun et al., 2017).

The method’s main assumptions are explicit. Discovery-at-random may fail if the probability that malware is labeled depends on features; misestimation of $U$9 can bias $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$0; and reliance on static permissions, APIs, and resolved IP endpoints leaves the system vulnerable to concept drift, obfuscation, or adversarial camouflage. This suggests that the Android AntiPure formulation is best interpreted as a training-set sanitation mechanism rather than a complete malware-analysis stack.

3. AntiPure in adversarial purification defenses

In a broader usage, AntiPure “refers broadly to defenses/attacks around adversarial purification,” and several systems in the supplied literature instantiate that broader purifier-centric design space (Lin et al., 2024). AToP, or “Adversarial Training on Purification,” defines the defended pipeline as

$$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$1

where $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$2 is a random transform, $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$3 is the purifier, and $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$4 is the fixed classifier. Its two key components are perturbation destruction by random transforms and adversarial fine-tuning of the purifier. Reported results include 72.85% robust accuracy for WideResNet-28-10 on CIFAR-10 under AutoAttack-$$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$5 at $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$6, and 76.37% for WideResNet-70-16, with better cross-threat generalization than AT-only baselines (Lin et al., 2024).

AGDM, or “Adversarial Guided Diffusion Models,” keeps the pretrained diffusion model fixed and adds robust latent guidance from an adversarially trained auxiliary network. Its guided DDPM step is written as

$$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$7

The supplied results state that AGDM improves robust accuracy by up to 7.30% on CIFAR-10, with PGD+EOT gains of +8.24% for WideResNet-28-10 and +9.53% for WideResNet-70-16 in the $$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$8 setting (Lin et al., 2024).

DiffAP reformulates diffusion purification around a random-sampling reverse process and mediator conditional guidance. The proposed reverse step is

$$\frac{t_m}{t_b}=\eta\cdot\frac{\#\text{benign}}{\#\text{malware}},\qquad \eta\ge 2,$$9

with $\eta=2$0. The reported outcome is a more than 20% robustness advantage under strong asynchronous PGD+EOT attacks, together with 10$\eta=2$1 sampling acceleration (Zhang et al., 2024).

NADD, or “Noise-Amplified Diffusion Defence,” pushes the same trajectory further by amplifying noise in both forward and reverse diffusion, then regularizing the reverse path with ring proximity correction. Its ImageNet result is 44.23% robust accuracy under AutoAttack with $\eta=2$2, an improvement of +2.07% over the previous best work, while reducing inference time to 1.08 seconds per sample with 29 reverse steps (Nguyen et al., 3 Jan 2026).

Taken together, these methods replace the earlier intuition that purification is merely denoising with a more explicit view of purifier design. Randomization, latent guidance, stochastic sampling, and geometric constraints all operate on the same underlying problem: removing perturbations without erasing semantics. A plausible implication is that modern AntiPure-style defenses are increasingly defined by how they control the purifier’s trajectory rather than by the generator class alone.

4. Task-specific purification systems in the broader AntiPure family

Several additional systems specialize purification for particular threat models. AMRM-Pure treats adversarial purification as preservation of semantic relationships among image patches. It defines the attention matrix

$\eta=2$3

and derives lower bounds linking attention matrix variation to adversarial reconstruction loss. Its robustly fine-tuned RAMRM-PureMaskDiT is reported to achieve 75.83% $\eta=2$4 robust accuracy on CIFAR-10 under AutoAttack and 36.87% on ImageNet, while outperforming DiffPure under BPDA+EOT20 (Dou et al., 5 Jul 2026).

SuperPure addresses localized and distributed adversarial patches by iterative downsampling, GAN-based super-resolution, and pixel-wise discrepancy masking. Its update rule is

$\eta=2$5

followed by an enhancement step

$\eta=2$6

The supplied evaluation states that SuperPure improves robustness against conventional localized patches by more than 20% on average, achieves 58% robustness against distributed patch attacks, and decreases defense end-to-end latency by over 98% compared to PatchCleanser (Khalili et al., 22 May 2025).

MalPurifier adapts adversarial purification to discrete Android feature vectors. Its Denoising AutoEncoder is trained with a dual-objective loss

$\eta=2$7

where $\eta=2$8 is input-space reconstruction loss and $\eta=2$9 aligns the purifier output with the detector’s internal representation. The reported system defends against 37 perturbation-based evasion attacks and consistently achieves robust accuracies above 90.91%, while remaining model-agnostic and plug-and-play (Zhou et al., 2023).

IMPure, or “Information Mask Purification,” argues that residual adversarial perturbations mainly come from same-position patches and similar patches. It reconstructs masked subsets in parallel and adds a random combination module,

$P$0

before computing perceptual loss. On ImageNet, the reported defended top-1 accuracies for Inception V3 are 84.28% under CW, 75.00% under PGD, and 74.94% under APGD-ce, with 87.30% clean accuracy (Liu et al., 2023).

These systems indicate that the broader AntiPure family is not tied to a single architecture. Diffusion models, mask autoencoders, super-resolution GANs, denoising autoencoders, and transformer-based reconstruction networks all appear, but each is specialized to a distinct perturbation model: norm-bounded image attacks, patch attacks, discrete malware evasion, or training-set contamination. This suggests that purification is a cross-domain design principle rather than a domain-specific algorithm.

5. AntiPure as a purification-resistant protective perturbation

The most direct use of the name AntiPure appears in the anti-purification literature on diffusion-based customization. There, AntiPure is introduced as a “simple diagnostic protective perturbation” within a purification–customization workflow in which a perturbed image $P$1 is first purified by $P$2 and then used by a downstream customization procedure $P$3. The paper formalizes an anti-purification task and adopts a practical objective that attacks purification itself rather than differentiating through the entire purification-and-customization pipeline (Yang et al., 17 Sep 2025).

Its perturbation generation begins from the diffusion noising equation

$P$4

with predicted denoised image

$P$5

Two guidance mechanisms define the method. Patch-wise Frequency Guidance computes a patch-wise DCT of $P$6 and emphasizes the bottom-right quadrant of each patch, interpreted as high-frequency content. Erroneous Timestep Guidance uses

$P$7

thereby encouraging confusion across timesteps in the purifier’s denoising schedule (Yang et al., 17 Sep 2025).

The full PGD objective is

$P$8

with projected updates under an $P$9 budget. The supplied implementation notes recommend $z(s)=1$0, $z(s)=1$1, $z(s)=1$2, and timestep sampling $z(s)=1$3 with $z(s)=1$4 for GrIDPure (Yang et al., 17 Sep 2025).

The reported evaluations use CelebA-HQ and VGGFace2, with DreamBooth and LoRA as downstream customization tasks. On DreamBooth with CelebA-HQ, AntiPure attains FID 81.15, ISM 0.6112, and BRISQUE 43.60; on DreamBooth with VGGFace2, it attains FID 90.77, ISM 0.5475, and BRISQUE 46.01. On LoRA, the corresponding FID values rise further to 109.63 on CelebA-HQ and 127.67 on VGGFace2. At the same time, AntiPure yields the lowest reported pre-purification perceptual discrepancy, with LPIPS 0.1392 and 0.2843 on CelebA-HQ and 0.1758 and 0.3884 on VGGFace2 for AlexNet- and VGG-based LPIPS, respectively (Yang et al., 17 Sep 2025).

The paper’s central claim is therefore not that purification can be strengthened, but that purification can be systematically stressed. AntiPure is framed as a diagnostic perturbation: it achieves minimal perceptual discrepancy and maximal distortion within the purification-customization workflow, exposing vulnerabilities in representative purification settings rather than merely attacking a downstream generator (Yang et al., 17 Sep 2025).

6. Critique, evaluation protocol, and the anti-purification turn

A major controversy in the purification literature concerns whether diffusion-based purification is robust at all. DiffBreak argues that adaptive, gradient-based attacks target the diffusion model rather than the classifier, causing purified outputs to align with adversarial distributions. It further attributes prior robustness claims to incorrect gradients and to an inappropriate single-purification evaluation protocol (Kassis et al., 2024).

The paper formalizes diffusion-based purification as a stochastic operator $z(s)=1$5 and studies attacks on

$z(s)=1$6

Its central theoretical claim is that, under exact differentiation and expectation over transformations, the attack changes the purifier’s path distribution rather than merely exploiting classifier sensitivity. To support this, DiffBreak introduces DiffGrad, described as the first reliable toolkit for differentiation through diffusion-based purification, and critiques common implementation errors involving adjoint timing, rounding, stochasticity reproduction, and guidance gradients (Kassis et al., 2024).

The paper also rejects the standard single-purification protocol because stochastic purification creates resubmission risk. It proposes a majority-vote alternative that aggregates predictions across multiple purified copies. Even under this stricter protocol, however, the reported robustness remains fragile. Under single purification, Full-DiffGrad reduces DiffPure robust accuracy on CIFAR-10 with WideResNet-28-10 to 8.59%. Majority vote yields only partial recovery; and the paper’s low-frequency systemic attack then drives robust accuracy under majority vote to approximately 0% on ImageNet and to 2.73–3.13% on CIFAR-10 for DiffPure, with similarly catastrophic results for GDMP (Kassis et al., 2024).

This critique changes the interpretive status of AntiPure-like defenses. Rather than assuming that purification inherently projects samples onto a natural manifold, the anti-purification literature treats the purifier itself as an attack surface. A plausible implication is that future AntiPure research will be judged less by clean denoising quality than by exact-gradient evaluation, stochastic-protocol validity, and resistance to adaptive attacks that optimize through the purifier’s internal dynamics.