In standard machine learning, missing data is usually treated as a nuisance—a broken sensor, a skipped survey question, or a corrupted file. We typically assume this missingness is "random" and use techniques like mean imputation or regression to fill in the blanks.
However, in high-stakes environments, missingness is often a deliberate choice. A loan applicant might hide a low credit score; a student might withhold a poor test result. Alternatively, a malicious actor might selectively delete data points from a training set to bias a model's behavior. In these cases, the missingness is not noise; it is a signal or a weapon.
The Imputation Vulnerability
The core problem lies in how we "fix" missing data. Traditional imputation methods assume that the missing values should look like the observed ones. If a model encounters a missing feature, it might fill it with the average value of that feature across the population.
As Anilesh K. Krishnaswamy, David Rein, and their colleagues demonstrate in Classification with Strategically Withheld Data, this creates a perverse incentive. If an individual knows their true value is below the population average, they can improve their predicted outcome by simply hiding that data. By treating missingness as a random occurrence, the model inadvertently rewards strategic concealment, leading to a significant drop in predictive accuracy.
Strategic Withholding at Inference
When an individual hides data to improve their own classification, they are engaging in strategic withholding. This is a game-theoretic challenge: the model designer wants to predict accurately, while the agent wants a favorable classification.
To counter this, Krishnaswamy et al. propose Incentive-Compatible (IC) models. The intuition is simple: the model should be designed so that no agent can improve their classification by hiding a feature. One way to achieve this is to "penalize" missingness—effectively assuming the worst-case value for any missing feature. If the model's response to a missing value is always less favorable than the response to the true value, the incentive to hide data disappears. Their work introduces algorithms like IC-LR (Incentive-Compatible Logistic Regression) and MINCUT to find the optimal balance between accuracy and incentive compatibility.
Adversarial Missingness at Training
While strategic withholding happens when a model is being used, a different threat emerges when a model is being trained. In Exploiting Missing Data Remediation Strategies using Adversarial Missingness Attacks, Alex Gittens, Moti Yung, and their co-authors introduce Adversarial Missingness (AM).
In this scenario, an attacker poisons a dataset not by adding "bad" data, but by selectively deleting "good" data. By using a bi-level optimization framework, an attacker can identify which specific data points to remove to shift the global statistics of the dataset.
Because standard pipelines use imputation to fill these gaps, the attacker can "steer" the imputation process. For example, by deleting high-value records for a specific feature, the attacker forces the mean-imputation of that feature to be lower, which in turn biases the learned weights of the final model. This attack is particularly dangerous because:
- It is stealthy: The attacker isn't injecting synthetic outliers; they are just making the data "disappear" in a way that looks like natural missingness.
- It bypasses defenses: Most data sanitization tools look for malicious inputs. They have no mechanism to detect the strategic absence of valid data.
Why This Works: Missingness as Information
The shared insight across this research is that missingness is informative.
In the strategic case, the fact that a student withheld a score suggests the score was likely low. In the adversarial case, the pattern of missingness is a signature of the attacker's goal.
Current machine learning security protocols are largely blind to this. By treating missing data as a statistical nuisance to be smoothed over, we create a "missingness channel" that agents and attackers can exploit. Moving forward, robust machine learning requires models that treat the presence or absence of data as a deliberate feature, rather than an accident of collection.