Skip to content

Bridging Models to Defend: A Population-Based Robust Adversarial Defense Strategy

Bridging Models to Defend: A Population-Based Strategy for Robust Adversarial Defense

arXiv· arXiv:2303.10225· Ren Wang, Yuxuan Li, Can Chen, Dakuo Wang, Jinjun Xiong, Pin-Yu Chen, Sijia Liu, Mohammad Shahidehpour, Alfred HeroDepartment of Electrical and Computer Engineering, Illinois Institute of Technology; School of Data Science and Society, University of North Carolina at Chapel Hill; Khoury College of Computer Sciences, Northeastern University; College of Arts, Media and Design, Northeastern University; Department of Computer Science and Engineering, University at Buffalo; IBM Thomas J. Watson Research Center; Department of Computer Science and Engineering, Michigan State University; Electrical Engineering and Computer Science Department, University of MichiganIndexed by this site · 23 days ago· Originally publishedDefenses

Translation in progress — untranslated text is omitted.

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors propose a population-based adversarial defense framework RMC/ERMC based on mode connectivity, achieving a maximum three-norm DLR of 53.97% on CIFAR-10 (ERMC-5, Table I).

Threat modelThe defender's goal is to make the model resist adversarial perturbations under multiple ℓp norms (p=1, 2, ∞) at inference time; the attacker uses PGD, Auto-Attack, and MSD to generate perturbations, with budgets δ of 8/255 (ℓ∞), 1 (ℓ2), and 12 (ℓ1) respectively; the victim system is an image classification neural network.

Experimental setupPreResNet110、WideResNet-28-10 et al. · CIFAR-10、CIFAR-100 et al. · Standard Accuracy、Robust Accuracy (ℓ∞/ℓ2/ℓ1-PGD) et al.
Evaluated models
PreResNet110WideResNet-28-10Vision Transformer-base
Benchmarks
CIFAR-10CIFAR-100ImageNet-100
Metrics
Standard AccuracyRobust Accuracy (ℓ∞/ℓ2/ℓ1-PGD)DLRUnionAA (ℓ∞/ℓ2/ℓ1)MSD
AI summaryThe researchers propose an adversarial defense framework based on robust mode connectivity (RMC), which uses two population learning stages to enhance the robustness of neural networks to diverse ℓp perturbations. The first stage searches the parameter space between two pretrained models for a highly robust continuous path, using the SRMC module to accelerate endpoint generation; the second stage combines RMC module optimization, with ERMC leveraging ℓ1 and ℓ∞ adversarially trained models to cover a broader range of p norms. Experiments on multiple datasets and architectures show that the method significantly improves robustness to ℓ∞, ℓ2, ℓ1, and mixed attacks.

The researchers propose an adversarial defense framework based on robust mode connectivity (RMC), which uses two population learning stages to enhance the robustness of neural networks to diverse ℓp perturbations. The first stage searches the parameter space between two pretrained models for a highly robust continuous path, using the SRMC module to accelerate endpoint generation; the second stage combines RMC module optimization, with ERMC leveraging ℓ1 and ℓ∞ adversarially trained models to cover a broader range of p norms. Experiments on multiple datasets and architectures show that the method significantly improves robustness to ℓ∞, ℓ2, ℓ1, and mixed attacks.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    How to make a neural network simultaneously resist adversarial perturbations of multiple ℓp norms (diversified ℓp robustness).

    Core problem: how to make a single neural network simultaneously robust to adversarial perturbations of multiple norms such as ℓ1, ℓ2, and ℓ∞ (the paper calls this diversified ℓp robustness, DLR).

    • Scenario and importance: Neural networks are used in safety-critical domains such as medical imaging, defect detection, and power systems, but are sensitive to adversarial perturbations imperceptible to the human eye; the authors consider this raises concerns about model trustworthiness.
    • Shortcomings of existing methods: The authors state that adversarial training (AT) and its variants are mostly designed for a single ℓp norm and lose robustness when encountering perturbations of other norms; existing multi-norm training methods (e.g., [13]–[18]) still rely on single-point parameter optimization and are prone to local minima or saddle points.
    • Core observation: Mode connectivity shows that low-loss paths exist between independently trained models, which can quickly generate a large number of candidate models; however, the authors use experiments to show that directly applying vanilla mode connectivity to two adversarially trained models does not yield high DLR (Fig. 3), and that adversarial robustness needs to be embedded into the path search.
    • Proposed in the paper: A two-stage population-based defense framework—Phase I's Robust Mode Connectivity (RMC, including the acceleration module SRMC), Phase II's RMC-based optimization, and ERMC for improved efficiency (including model ensemble).
    • Threat model:
      • Attacker goal: At inference time, use ℓp-norm-bounded perturbations to make the classification model err; the experiments focus on p = 1, 2, ∞.
      • Attacker knowledge/capability: Use PGD, Auto-Attack (AA), and MSD to generate perturbations; perturbation budgets δ = 8/255 (ℓ∞), 1 (ℓ2), 12 (ℓ1).
      • Victim system: Image classification neural networks (PreResNet110, WideResNet-28-10, Vision Transformer-base).
      • Defender setting: Defense is performed at the training stage, and multiple adversarially trained models can be used for population optimization.
  2. What related research exists?

    Related work falls into four groups: adversarial attacks, adversarial training, multi-norm defenses, and population-based neural network learning.

    Adversarial attacks

    • FGSM [6] and PGD [8]: construct attacks using the local gradients of the target model; output diversified sampling [25] improves initialization. The paper states that these methods may give inaccurate robustness measurements due to improper hyperparameter tuning and gradient masking.
    • Auto Attack [26]: combines four attacks and adjusts step sizes; MSD [16]: incorporates multiple perturbation models into the steepest descent projection at each step. This paper uses PGD, AA, and MSD as evaluation attacks.

    Adversarial training defenses

    • AT [8]: pioneered adversarial training with min-max optimization; subsequent work includes TRADES [10] (robustness-accuracy trade-off), dynamic adversarial training [27], and semi-supervised robust training [28].
    • Fast AT: [11], [12], [29], [30] propose faster approximate AT algorithms. The paper points out that a common problem of these methods is that training targets only a single ℓp norm, and robustness drops substantially when encountering other perturbation types.

    Multi-norm (diversified ℓp) defenses

    • [14]: the paper states it is the only work providing a certified defense; [15]: improves resistance to stronger attacks by withholding specific inputs.
    • [13]: the internal loss takes the maximum or average loss across perturbation types; E-AT [17]: uses a fine-tuning strategy; MSD Defense [16]: incorporates multiple perturbation models at each step. The paper points out that these works still optimize a single set of parameters, and the deficiency in multi-norm robustness remains unresolved; this paper approaches it from the perspective of population-based optimization.

    Population-based neural network learning

    • [31]: trains multiple model instances in parallel and selects the best to reproduce; [32]: evolutionary stochastic gradient descent. The paper states that such methods learn slowly and ignore adversarial robustness.
    • [33]: mimics the immune system to improve robustness at inference time; mode connectivity [23] can be viewed as faster population-based learning; [34] analyzes mode connectivity under backdoors/error injection/single perturbation, and this paper states that it extends this to robust population optimization over multiple ℓp perturbations.

    Baselines and benchmarks: the comparison baselines are ℓ∞-AT [8], RMC-RI (RMC on two randomly initialized models), E-AT [17], and MSD Defense [16]; the datasets are CIFAR-10, CIFAR-100, and ImageNet-100.

    Positioning of this paper: the authors state that this paper unifies and extends their workshop paper [1] (Phase I path discovery) and Deep Adversarial Defense [2] (preliminary validation of ERMC) into a complete two-stage framework, and adds SRMC, theoretical guarantees, and more experiments.

  3. How does the paper address this problem?

    Two-stage population-based framework: RMC searches for robust paths, Phase II combines RMC units and selects the best, and ERMC improves efficiency.

    Overall idea: Combine mode connectivity with multi-norm adversarial training, search for a continuous path between two (or more) adversarially trained models on which every point is a highly robust model, and then use RMC as the basic unit for population-based optimization and selection. The framework is named Robust Mode Connectivity (RMC)-oriented adversarial defense.

    Problem setting and metric

    • Diversified robustness (DLR): Defined as the model's robust accuracy under the worst perturbation type, i.e., mini∈[S] (∑(x′,y)∈D̂i 1f(x′,θ)=y)/(|D̂i|), where D̂i is the adversarial data generated by the i-th ℓp perturbation; the paper notes that a worst-case sample-wise definition can also be used, which essentially measures the same quantity.
    • Perturbation setting: The experiments use ℓ∞, ℓ2, and ℓ1 norms, with δ of 8/255, 1, and 12, respectively.

    Phase I: Robust Mode Connectivity (RMC)

    • Motivating experiment: Directly applying vanilla mode connectivity to ℓ∞-AT and ℓ2-AT models (PreResNet110, each trained for 150 epochs) (retrained for 50 epochs) yields high loss along the path and low robust accuracy under both attack types (Fig. 3); based on this, the authors argue that adversarial robustness needs to be embedded into the path search.
    • RMC objective: Add inner maximization to the mode connectivity objective, minθ Et∼ U(0,1) E(x,y)∼ D0 ∑i∈ I maxDisti(x′,x)≤δi L(ϕ(t;θ);(x′,y)); the curve is fixed as the quadratic Bezier curve ϕ(t; θ) = (1−t)²θ₁ + 2t(1−t)θ + t²θ₂.
    • Solver: The inner layer uses Multi Steepest Descent (MSD), considering all perturbation types simultaneously at each step (Algorithm 1); the authors state that since the number of perturbation types I is constant (2 or 3), the complexity order is the same as conventional AT.
    • Theoretical motivation (Theorem 1): Under assumptions such as independent and identically distributed peak epochs for each data point-perturbation type pair, gives a formula for the number of models K required to achieve "alignment" at least once; the authors note that the independence assumption may not hold in practice, and that the purpose is only to illustrate the inherent advantage of the population method over a single model.
    • SRMC acceleration module: First train an ℓp-AT model ϕ(0; θ), then fine-tune it for a few epochs with another perturbation Distj to obtain the other endpoint ϕ(1; θ), avoiding training both endpoints from scratch.

    Phase II: RMC-based optimization

    • Procedure (Algorithm 2, three perturbation types example): 1) Train three models with ℓ1-, ℓ2-, and ℓ∞-AT for T epochs each; 2) Use RMC to connect the ℓ2-AT and ℓ1-AT models and the ℓ∞-AT and ℓ1-AT models, obtaining two trajectories; 3) Randomly sample points in the optimal region of each trajectory; 4) From these two points, train with ℓ∞-AT and ℓ2-AT for T epochs respectively to obtain new endpoints; 5) Use RMC again (I containing three perturbation types) to connect them, and take the model at the point topt with the highest DLR.
    • Design rationale: The authors state that randomly sampling points near the highest-DLR point is to increase diversity; parameter curves derived from different models can be computed in parallel, and when parallelized, the execution time equals the time to generate one curve.

    ERMC and model ensemble

    • Theoretical basis: Cites the theorem in [14]—if an affine classifier (including ReLU CNN) is robust to ℓ1 and ℓ∞ perturbations, then it has satisfactory robustness to ℓp perturbations for 1 ≤ p ≤ ∞ (cited in the paper as Theorem 2).
    • ERMC (Algorithm 3): The endpoint ϕ(0; θ) is trained with ℓ∞-AT, copied and fine-tuned with ℓ1-AT for 10 epochs to obtain ϕ(1; θ), and then the RMC objective containing both ℓ1 and ℓ∞ perturbations is optimized; the authors state that ERMC requires only one connection, and is therefore efficient.
    • ERMC-n ensemble: On the path, take the interval [a, b] satisfying the thresholds α∞ (ℓ∞ robust accuracy) and α1 (ℓ1 robust accuracy), select n models at intervals t = a + (b−a)/(n−1)·i, and average the final-layer outputs; when there are multiple non-adjacent qualifying intervals, allocate slots in proportion to interval length.
  4. What experiments did the paper conduct?

    Evaluated on CIFAR-10/100, ImageNet-100, and three architectures, ERMC-5 achieves a DLR of 53.97% under three perturbation types.

    Experimental setup

    • Subjects under test: the defense methods are installed on three architectures: PreResNet110, WideResNet-28-10, and Vision Transformer-base; the default configuration is CIFAR-10 + PreResNet110.
    • Datasets: CIFAR-10, CIFAR-100, ImageNet-100.
    • Baselines: standard ℓ∞-AT [8], RMC-RI (RMC on two randomly initialized models), E-AT [17], MSD Defense [16].
    • Attacks and perturbation budgets: ℓ∞/ℓ2/ℓ1-PGD, Auto-Attack (AA), MSD attack; δ = 8/255 (ℓ∞), 1 (ℓ2), 12 (ℓ1).
    • Metrics: clean accuracy (Standard Accuracy), robust accuracy under each PGD attack, Union (worst-case sample-wise accuracy under the three PGD attacks), AA (ℓ∞/ℓ2/ℓ1), MSD attack accuracy, DLR (robust accuracy under the worst perturbation type).
    • Training amount and hardware: endpoint models are trained for 50/150 epochs, and the path is trained for an additional 50 epochs; all Phase II experiments train for 200 epochs (parallelism allowed); two NVIDIA RTX A100 GPUs. The paper does not report the number of repetitions or variance.

    Main results

    Table I (CIFAR-10 + PreResNet110; the DLR in the "two perturbation types" rows is taken only over ℓ∞/ℓ2-PGD, and the "three perturbation types" rows over the three PGD attacks):

    Swipe horizontally to read the table

    MethodStandardℓ∞-PGDℓ2-PGDℓ1-PGDDLRUnionMSD
    ℓ∞-AT85.00%49.03%29.66%16.61%/21.85%15.27%
    MSD-Defense (two types)81.61%48.57%45.92%35.64%45.92%34.37%45.72%
    RMC-RI (two types)63.08%37.4%38.44%30.47%30.47%29.22%36.9%
    RMC (two types)80.90%48.19%48.63%38.05%48.19%36.3%46.52%
    RMC-based opt (two types)81.36%48.89%49.03%38.83%48.89%36.86%47.18%
    MSD Defense (three types)81.35%40.14%48.58%47.50%40.14%38.35%38.20%
    E-AT (three types)79.3%44.07%49.12%49.82%44.07%41.08%42.67%
    RMC-based opt (three types)81.76%46.21%51.86%46.23%46.21%41.47%44.75%

    (Table I also has rows for the 5-model ensemble, SRMC variants, and ERMC-1/3/5, see below.)

    • The authors state: under two perturbation types, the DLR of RMC is 2.27% higher than MSD-Defense and 10.79% higher than RMC-RI; under three perturbation types, the DLR of RMC-based optimization is 6.07% higher than MSD Defense and 2.14% higher than E-AT, and the Union/AA/MSD metrics are also higher.
    • The authors point out that RMC-based optimization has a trade-off between clean accuracy and DLR: its clean accuracy is 3.24% lower than ℓ∞-AT, but the decrease is smaller than that of MSD Defense and E-AT; current model selection prioritizes DLR.
    • The authors state that RMC-based optimization with the SRMC module (three types) has a DLR of 46.10%, similar to (slightly lower than) the version without SRMC, with slightly higher AA and Union (Table I).

    ERMC and ensembles

    • ERMC-n (Table I, three perturbation types): ERMC-1 has a DLR of 46.54%, which the authors state is similar to RMC-based optimization; ERMC-3 has a DLR of 48.49%; ERMC-5 has a DLR of 53.97%, Union 51.41%, and MSD 49.78%; the authors state that performance improves as n increases, and ERMC-5 outperforms all other methods on DLR, Union, and MSD.
    • 5-model ensemble (Table I): 5 models selected from the curves in Fig. 9 are ensembled (thresholds α∞ = 37%, α1 = 43%), with DLR 51.05% and Union 49.39%; the authors state that ensembling can further improve RMC-based optimization.
    • ERMC path shape (Fig. 10): connecting ℓ∞ and ℓ1 models on CIFAR-10, CIFAR-100, WideResNet-28-10, and ViT-base all yields highly robust paths; the authors observe that endpoint models are robust to the perturbation type they were trained on and relatively fragile to the other.

    Path search and efficiency

    • RMC paths (Fig. 4): between ℓ∞-AT and ℓ2-AT endpoints with the MSD inner solver, the paths at 50/100/150 epochs all contain high-DLR points, with the highest DLR in panel (a) being 48.19%; more epochs make the path smoother, and the optimal-point DLR is similar across the three epoch settings, on which basis the authors conclude that selecting the best model requires only a small number of epochs.
    • Cross-dataset/architecture (Fig. 5): RMC finds paths containing high-DLR points on CIFAR-100, ImageNet-100, WideResNet-28-10, and ViT-base (two perturbation types).
    • SRMC efficiency (Section VI-C): on CIFAR-10 + PreResNet110 on a single GPU, the full RMC pipeline (150 epochs for each of the two endpoints + 50 epochs for the curve) averages 2750 minutes, versus 1780 minutes for SRMC; on ViT on a single GPU, RMC averages 5785 minutes and SRMC 3592 minutes.

    Other ablations and analyses

    • Two perturbation types, a single intermediate optimal point (sampled at t ∈ [0.77, 0.83]): the optimal-point DLR is 48.8% (t = 0.72, Fig. 7).
    • Two perturbation types, two intermediate optimal points (t ∈ [0.27, 0.33] and [0.77, 0.83]): the optimal DLR is 48.89% (t = 0.71, Fig. 8); the authors state that the accuracy drop around t ≈ 0.5 stems from RMC using only 50 epochs.
    • Three perturbation types, Phase II (T = 50, RMC with an additional 50 epochs): the optimal-point DLR is 46.21% (t = 0.93, Fig. 9).
    • Exception: in Table I, ERMC-3's AA-ℓ∞ is 42.88% (its lowest AA term), lower than ERMC-1's 43.97%; the authors do not explain this.
  5. What are some points that can be further explored?

    The authors mention limitations such as theoretical assumptions and the accuracy-robustness trade-off; future directions include foundation models and certified robustness.

    Limitations and future directions noted by the authors

    • Theoretical assumptions: The independence assumption of Theorem 1 "may not hold in practice"; the authors state that its purpose is only to show the inherent advantage of the population method over single-model solutions (Section IV-A3).
    • Accuracy-robustness trade-off: The authors state that the current RMC model selection prioritizes DLR over clean accuracy, and that future implementations can adjust this to achieve a better balance (Section VI-F, observation (4)).
    • Future directions: The authors state that future work could extend RMC to large-scale foundation models, combine it with certified robustness techniques, and explore applications in safety-critical domains such as healthcare and power systems (Section VII).

    Scope of experimental coverage

    • The architectures tested are PreResNet110, WideResNet-28-10, and Vision Transformer-base, 3 in total; the datasets are CIFAR-10, CIFAR-100, and ImageNet-100, 3 in total (Section VI-A).
    • The quantitative comparison table (Table I) is given only on CIFAR-10 + PreResNet110; CIFAR-100, ImageNet-100, and other architectures are shown only as path plots (Fig. 5, Fig. 10).
    • The perturbation types are ℓ∞, ℓ2, and ℓ1, with δ of 8/255, 1, and 12, respectively; the evaluated attacks are PGD, Auto-Attack, and MSD (Section VI-A).
    • The efficiency data are single-GPU minutes: on PreResNet110, RMC 2750 minutes vs SRMC 1780 minutes; on ViT, 5785 vs 3592 minutes (Section VI-C).
    • The paper does not report the number of experimental repetitions, variance, or statistical significance tests; Theorem 1 gives a formula for population size but no corresponding empirical validation is performed.
  6. Summarize the main content of the paper

    Population-based mode connectivity defense framework, with a maximum three-norm DLR of 53.97% (ERMC-5, Table I).

    An adversarial robustness defense paper that addresses the problem that single-ℓp-norm adversarial training cannot cover multiple perturbation types, and proposes the RMC-oriented adversarial defense framework based on population-based optimization.

    • Method: Phase I's RMC embeds the inner maximization of adversarial training into mode connectivity, searching for a highly robust path between two AT models of different norms (solved by MSD, Bezier curve); the SRMC module uses fine-tuning instead of training the second endpoint from scratch to accelerate; Phase II uses RMC as the basic unit for multi-generation population optimization and selection; ERMC, based on the theorem in [14], connects the ℓ∞ and ℓ1 models only once, and uses ERMC-n to ensemble n models on the path.
    • Key results (CIFAR-10 + PreResNet110, Table I): Under two perturbation types, RMC achieves a DLR of 48.19%, higher than MSD-Defense's 45.92%; under three perturbation types, RMC-based optimization achieves a DLR of 46.21%, higher than MSD Defense's 40.14% and E-AT's 44.07%; ERMC-5 achieves a DLR of 53.97% and Union of 51.41%, and the authors state that it outperforms all compared methods on DLR, Union, and MSD.
    • Generalization and efficiency: RMC finds high-DLR paths on CIFAR-100, ImageNet-100, WideResNet-28-10, and ViT-base (Fig. 5, Fig. 10); SRMC reduces training time on PreResNet110 from 2750 minutes to 1780 minutes (single GPU).
    • Authors' conclusion: The authors state that the framework establishes population-based mode connectivity as a generalizable adversarial defense paradigm, while acknowledging a trade-off between clean accuracy and DLR (RMC-based optimization's clean accuracy is 3.24% lower than ℓ∞-AT), and plan to extend it to foundation models and combine it with certified robustness in the future.

Related papers

Show 3 more papers
Read originalarxiv.org