Initialization Sensitivity in Adversarial Optimization

21 concepts16 papers1 proposal

In standard deep learning, initialization is often viewed as a technical necessity to ensure convergence. However, in the context of adversarial machine learning, the "starting point" of an optimization process—whether it is the initial weights of a model or the initial noise added to an image—dictates the ultimate security and effectiveness of the system. Recent research demonstrates that the adversarial loss landscape is significantly more jagged and prone to sharp minima than standard landscapes, making the choice of initialization a primary determinant of success for both attackers and defenders.

The Adversarial Landscape Bottleneck

Adversarial training (AT) is the standard for building robust models, but it suffers from optimization instabilities. Liu et al. identify that the adversarial loss landscape is characterized by extreme jaggedness. They argue that applying a large adversarial perturbation budget at the very beginning of training can trap a model in suboptimal regions near its random initialization. Because the model is untrained, overwhelming perturbations prevent it from extracting meaningful features, leading to "collapsed" solutions. To counter this, the authors propose Periodic Adversarial Scheduling (PAS), which cycles the attack budget to help the model navigate this difficult landscape.

The sensitivity extends to the model's weights themselves. Bergou et al. demonstrate that weight initialization and training duration are as critical for adversarial defense as sophisticated pre-processing. A model's long-term security profile is effectively "seeded" by its initial parameter state, a factor often overlooked in favor of complex architectural changes.

Finding the Decision Boundary

For attackers, the challenge is finding the smallest possible perturbation that causes misclassification. Many high-performance attacks, such as those that "walk" along the decision boundary, require a valid adversarial starting point to begin their search. Brendel et al. show that the quality and convergence of these boundary-following attacks are heavily influenced by this initial point. If the starting adversarial example is poor, the optimization may never reach a query-efficient or low-norm solution.

This requirement is particularly acute in sparse attacks. Cinà et al. note that many $\ell_0$ attacks must be seeded with an image that is already misclassified (such as pure noise) and then work backward. If a suitable starting point cannot be found, the attack fails entirely. Similarly, Pooladian et al. argue that simple noise-based initialization is often insufficient for complex models, necessitating more robust methods to find the "wrong side" of the decision boundary.

Initialization as a Tool for Transferability

One of the most significant breakthroughs in this area is using initialization to improve adversarial transferability—the ability of an attack crafted on one model to succeed on another. Liu et al. treat transferability as a bilevel optimization problem where the "upper level" specifically optimizes the initial perturbation point. By finding a starting point that avoids local minima specific to a single surrogate model, the attack follows an optimization path that leads to regions of the loss landscape shared across diverse architectures.

Zhang et al. further this by introducing Randomized Global Initialization (RGI). Instead of starting from a single point, they average gradients from multiple perturbed versions of an input. This provides a "global guidance" that captures a broader view of the loss landscape, preventing the attack from getting stuck in narrow, non-transferable optima.

Efficiency through Informed Starting Points

In black-box settings, where every attempt costs a "query" to the target model, initialization is a matter of efficiency. Xiang et al. show that selecting initial samples closer to the decision boundary with smaller norms can drastically reduce the number of queries needed for convergence. Even simple heuristic strategies can yield massive gains; Lapid et al. found that a "vertical stripes" initialization was surprisingly effective at reducing the query budget for evolutionary attacks.

For defenders, "priming" the optimization can also lead to faster training. Jia et al. introduced FGSM-SDI, which replaces random noise initialization with a learnable, sample-dependent starting point. By using a lightweight network to predict the best place to start an attack during training, they achieve the robustness of expensive multi-step methods at a fraction of the computational cost.

Risks in Multi-Agent and Federated Systems

The sensitivity to starting conditions becomes a systemic risk in distributed and evolving environments. In federated learning, Singh and Vaswani highlight that the initialization phase is the most vulnerable to Byzantine (malicious) actors. A single biased starting point can lead the entire global optimization astray.

In the realm of Large Language Models (LLMs), Wen et al. show that in co-evolving attacker-defender games, the initial capabilities of the attacker determine the final safety of the model. If the attacker starts too weak, the defender never learns to resist complex, "long-tail" vulnerabilities. Similarly, Tang et al. demonstrate that the success of multi-turn jailbreaking is fundamentally dependent on the quality of the initial attack trajectory, which sets the stage for all subsequent dialogue fabrication.

Key Foundational Work

Accurate, reliable and fast robustness evaluation (2019) — contributed 2 concept elements to this direction.
The LogBarrier adversarial attack: making effective use of decision boundary information (2019) — contributed 1 concept elements to this direction.
On the Loss Landscape of Adversarial Training: Identifying Challenges and How to Overcome Them (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_adversarial_initialization_sensitivity_3e2b58e1,
  title = {adversarial initialization sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-initialization-sensitivity/rdr_3e2b58e1737940499f0b165ca2b7230b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_adversarial_initialization_sensitivity_3e2b58e1_v1,
  title = {adversarial initialization sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-initialization-sensitivity/rdr_3e2b58e1737940499f0b165ca2b7230b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake