Strategic Sample Selection for Efficient Data Poisoning

This direction is merged into another direction: backdoor attacks. Its last version stays readable and citable here.

Version 1 of 131 papers39 conceptsfrozen 2026-10-04

In traditional backdoor attacks, an adversary injects a "trigger" into a random subset of training data to force a specific misclassification at inference time. However, recent research has shifted focus from what the trigger looks like to where it is placed. It turns out that not all training samples are equally susceptible to poisoning. By strategically selecting high-influence data points, attackers can achieve near-perfect success rates while corrupting less than 0.1% of the dataset, making the attack significantly stealthier and harder to detect.

The Intuition of Leverage

To understand why selection matters, think of a machine learning model’s decision boundary as a flexible sheet held in place by training data points. If you want to warp that sheet to create a "pocket" (a backdoor) around a specific trigger, you could try to push on every point randomly. But it is far more efficient to find the specific points that act as "anchors" for the boundary.

As Hossein Souri and Micah Goldblum demonstrate in Sleeper Agent, samples with a high gradient norm are the most influential. These points cause the largest updates to the model's parameters. By perturbing these specific samples, an attacker exerts disproportionate control over the model's learning process.

Targeting the Model’s Weak Points

Beyond simple gradient magnitude, researchers have identified several specific "vulnerabilities" in the data distribution that make certain samples ideal for poisoning:

  • Forgettable Samples: Pengfei Xia and Ziqiang Li introduce the "Filtering-and-Updating Strategy" (FUS) in Data-Efficient Backdoor Attacks. They target "forgettable" samples—data points that the model learns and then forgets multiple times during training. Because the model already struggles with these points, they are easier to "hijack" with a backdoor trigger without damaging the model's overall accuracy.
  • Boundary Samples: In Confidence-driven Sampling for Backdoor Attacks, Han Xu and colleagues argue that poisoning samples near the model's decision boundary (where the model is most uncertain) is more effective. These points are naturally "unstable," allowing a small malicious perturbation to shift them across the boundary more easily than a high-confidence "core" sample.
  • Representation Distance: XINGSHUO HAN proposes Representation Distance (RD) scoring in Computation and Data Efficient Backdoor Attacks. This method identifies samples that are mathematically "far" from the target class in the model's internal feature space. Forcing these distant samples into the target class creates a stronger gradient signal, effectively "stretching" the model's logic to include the backdoor.

Mathematical Optimization of Poisoning

Rather than using heuristics, some researchers treat sample selection as a formal optimization problem. In Game of Trojans, Arezoo Rajabi and Bo Li leverage submodularity to greedily select the minimum set of points required to embed a backdoor. This mathematical framework proves that an adaptive adversary can bypass statistical detectors with 100% certainty by minimizing their poisoning footprint.

Similarly, Arun Suggala and Sewoong Oh show in Label Poisoning is All You Need that influence functions—a tool from robust statistics—can rank which clean images would be most "damaging" if their labels were flipped. This allows for successful attacks by corrupting only the labels of a tiny fraction of the dataset, leaving the images themselves untouched.

Efficiency as a Path to Stealth

The primary motivation for these selection strategies is stealth. Traditional attacks that poison 5-10% of a dataset create statistical anomalies that automated data-sanitization tools can catch. However, when the poisoning ratio drops to 0.1% or lower, the malicious signal is buried in the natural noise of the dataset.

For instance, Zihao Zhu demonstrates in Boosting Backdoor Attack with A Learnable Poisoning Sample Selection Strategy that optimizing the selection of less than 0.3% of training data can increase attack success from negligible levels to over 36%. In high-stakes scenarios like Off-Policy Evaluation, Elita Lobo shows that corrupting just 3% of high-influence transitions can lead to value estimate errors exceeding 300%.

Forensic Implications

The same principles used to attack models can also be used to defend them. If we know that attackers must target high-influence samples, we can build forensic tools to trace the source of a poison. Evan Rose and Alina Oprea developed UTrace, which uses gradient similarity to identify which specific data owner in a collaborative learning environment contributed the "top-k" gradients responsible for a model's misbehavior. By understanding the "leverage" of specific samples, researchers are moving toward a more precise, surgical understanding of both how to break and how to fix machine learning pipelines.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_strategic_sample_selection_for_efficient_f6b9f5f9,
  title = {Strategic Sample Selection for Efficient Data Poisoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strategic-sample-selection-for-efficient-data-poisoning/rdr_f6b9f5f95ed345a19e0c2d48935c3016},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_strategic_sample_selection_for_efficient_f6b9f5f9_v1,
  title = {Strategic Sample Selection for Efficient Data Poisoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strategic-sample-selection-for-efficient-data-poisoning/rdr_f6b9f5f95ed345a19e0c2d48935c3016/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}