Black-Box Adversarial Example Attacks on Machine Learning based Election Systems

Subek Acharya, Syed Ayon, Khaleque Md Aashiq Kamal, Alina Barnett, Surya Eada, Kaleel MahmoodarXiv 2026 · arXiv

TL;DR

This paper demonstrates that machine learning models used for ballot classification are vulnerable to realistic "black-box" adversarial attacks, where the attacker has no knowledge of the model's internal parameters. By physically printing and scanning 20,000 ballots, the researchers reveal a significant robustness gap between digital and physical domains, proving that while physical distortion offers some protection, it is not a complete defense against sophisticated election tampering.

What Problem Does This Paper Address?

As election systems move toward automation, machine learning (ML) is increasingly proposed for ballot classification—specifically, determining whether a voter has filled in a bubble or left it blank. While these models are highly accurate, they are susceptible to adversarial examples: inputs designed to trick a classifier into making a wrong prediction through subtle, often imperceptible, perturbations.

Previous research in this niche has focused on "white-box" attacks, where the attacker has full access to the model's architecture and weights. This is an unrealistic assumption for proprietary election systems. This paper tackles the more dangerous and realistic black-box threat model, where an attacker can only query the system to see its output (e.g., "blank" or "filled") or its confidence scores. The goal is to understand if an attacker with a compromised printer could systematically flip an election by adding invisible noise to ballots.

Figure 2: An attack on an election system using adversarial examples. Under this attack scenario three outcomes are pos- sible. In two of the cases, a vote is given to the attacker’s pre- ferred candidate (Candidate A). In one case, neither candidate receives a vote. The main focus of this work is determining which black-box adversarial example method is most effective when physically realized, for an election attack.

Methods

The authors evaluated 11 different black-box attacks across three categories:

  • Decision-Based Attacks: The attacker only sees the final class label (e.g., ADBA, RayS, SurFree).
  • Score-Based Attacks: The attacker sees class probabilities, allowing for more guided perturbations (e.g., Square Attack, SA-MOO).
  • Transfer-Based Attacks: The attacker trains a "synthetic" surrogate model to mimic the target and generates attacks against it, hoping they "transfer" to the real system.

A key contribution is the physical attack pipeline. Instead of just simulating attacks digitally, the team generated adversarial examples, printed them using a standard laser printer, and scanned them back in using a high-fidelity scanner. They then used a UNet autoencoder to denoise the scans before classification, mimicking a real-world election processing workflow.

Figure 1: End-to-end physical attack pipeline. Adversarial examples are generated using a black-box attack method, printed, scanned and converted into a PyTorch tensor for model prediction. This setup models the attack process that would hypotheti- cally be used during a U.S. election to alter the outcome in favor of one candidate.

Datasets

The study utilized the UConn Bubbles with Marginal Marks dataset, which consists of 40x50 grayscale images of ballot bubbles. It includes two subsets:

  • Bubbles: Only clean blank and filled bubbles.
  • Combined: Includes "marginal marks" like pen rests, checkmarks, and scribbles to better reflect real-world voter behavior.

The researchers trained eight diverse models, ranging from traditional SVMs to modern Vision Transformers (CaiT) and even Spiking Neural Networks (SNNs).

Key Insights and Ideas

  • The Physical-Digital Robustness Gap: Adversarial noise is incredibly fragile. The "noise" introduced by the physical process of printing and scanning (ink bleed, paper texture, sensor noise) often destroys the carefully crafted adversarial signal. For example, the Square Attack's effectiveness dropped by nearly 40% when moving from digital to physical.
  • Transferability is the Real Threat: Even when a model appeared perfectly robust against direct physical attacks, it often failed when presented with adversarial examples generated for other models. This is the first transferability study in the voting domain, proving that "security through obscurity" (not knowing the model) is insufficient.
  • Architecture Matters: Vision Transformers (CaiT) and Spiking Neural Networks showed different levels of vulnerability compared to standard CNNs, suggesting that model ensemble defenses (like the "Barrier Zones" tested) are necessary.
  • Imperceptibility: The attacks were tuned to be "imperceivable" based on thresholds established by election security experts, ensuring the tampering wouldn't be caught by a human eye during a manual audit.
Figure 3: Digital examples with varying perturbation values for all attacks on the CaiT model. Corresponding numerical results are given in Table 6.

Key Results

  • Digital Dominance: In the digital domain, the Square Attack was the most lethal, reducing the robust accuracy of five out of eight models to near zero.
  • Physical Resilience: In physical tests, robustness increased significantly across the board. However, no model was perfectly secure. The ADBA attack remained effective against several models even after printing and scanning.
  • Defense Performance: The authors tested Adversarial Training (AT) and Barrier Zones (BARZ). While both improved robustness, neither provided a "silver bullet" solution, as some transfer attacks could still bypass these defenses.

Limitations and Future Work

The primary limitation acknowledged is the assumption that an attacker has access to a compromised printer within the election pipeline. While a high bar, it is a critical vulnerability to consider.

Future work includes:

  • Exploring attacks on more complex ballot types beyond simple bubbles.
  • Developing more robust denoising techniques that don't inadvertently help the attacker.
  • Investigating "universal" adversarial perturbations that could be applied to all ballots regardless of the specific voter's marks.

Summary updated

Abstract

Machine learning models have shown exceptional accuracy for ballot classification in election systems. Recent research has demonstrated that these ballot classification models are vulnerable to adversarial example attacks. This is a critical security issue because a successful attack could alter the outcome of an election and undermine the democratic process. However, all adversarial example attacks in the voting domain have assumed a white-box threat model, where the classifier parameters are known. In this work, we advance the field of machine learning in election systems by analyzing a realistic black-box adversary. We implement and evaluate eleven state-of-the-art black-box attacks: query based attacks ADBA, Square, SurFree, RayS, and SA-MOO. We also test adaptive transfer based black-box attacks with APGD-$l_\infty$, APGD-$l_2$, APGD-$l_1$, PGD $l_0$, PGD $l_0$+$l_\infty$, and PGD $l_0$+$σ$. Our attacks span multiple different norms ($l_\infty$, $l_2$, $l_1$, and $l_0$) and three different black-box adversarial threat models (decision based, score based and transfer based). In addition to digital analyses, we physically print and scan $20,000$ adversarial examples to evaluate the real-world performance of different black-box attacks. Our finding demonstrate an even larger physical-digital robustness gap exists for black-box adversarial examples than for white-box adversarial examples. Lastly, we conduct the first ever adversarial example transferability study in the voting domain and test adversarial training and barrier zones as two possible black-box defense strategies.
BibTeX
@misc{acharya2026black,
  author = {Subek Acharya and Syed Ayon and Khaleque Md Aashiq Kamal and Alina Barnett and Surya Eada and Kaleel Mahmood},
  title = {{Black-Box Adversarial Example Attacks on Machine Learning based Election Systems}},
  year = {2026},
  eprint = {2611.02166},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02166},
}
Figure