Equal Error Rate (EER)

The EER represents the operating point where security and convenience are balanced. A lower EER indicates higher overall system accuracy.

• At the EER, the probability of accepting an impostor equals the probability of rejecting a genuine user.
• It is derived from the intersection of the FAR and FRR curves plotted against the decision threshold.
• In RF fingerprinting, this threshold controls the similarity score required to authenticate a transmitter.

The EER is visualized on a DET curve, a modified ROC curve that plots FRR against FAR on a normal deviate scale.

• The EER is the point where the DET curve intersects the diagonal line y=x.
• This visualization makes it easy to compare the performance of different fingerprinting algorithms.
• A curve pushed closer to the origin indicates superior discrimination between authorized and unauthorized devices.

Unlike FAR and FRR which vary with the system's decision threshold, the EER is a threshold-independent metric.

• This makes it ideal for comparing the intrinsic discriminative power of different feature extractors or neural network architectures.
• It answers the question: 'How good is this model at separating classes, regardless of how I set the sensitivity?'
• For few-shot device enrollment, a low EER confirms that the learned embedding space effectively separates known devices from impostors.

EER is computed by sweeping the decision threshold and finding the value where FAR and FRR intersect.

• FAR = False Positives / (False Positives + True Negatives)
• FRR = False Negatives / (False Negatives + True Positives)
• In continuous systems, the EER is often found by interpolating between the two closest operating points where FAR > FRR and FRR > FAR.
• A common target for high-security RF authentication is an EER below 1%.

The EER is directly related to the Area Under the ROC Curve (AUC). A system with a higher AUC will have a lower EER.

• The ROC curve plots True Acceptance Rate (1-FRR) against FAR.
• While the ROC curve shows performance across all thresholds, the EER summarizes it into a single actionable number.
• For engineers, the EER is often more intuitive than AUC for setting concrete security requirements.

The EER assumes a closed-set identification problem where every probe belongs to a known class. In open set emitter recognition, its utility is limited.

• EER does not measure the system's ability to reject unknown, never-before-seen transmitters.
• For open set problems, metrics like the Open Set Classification Rate or detection error trade-off curves at specific false alarm rates are more informative.
• In few-shot enrollment, the EER is best used to evaluate the core embedding space before adding an open set rejection layer.

The probability that a biometric system incorrectly authenticates an unauthorized user or device. FAR represents a critical security breach—granting access to an impostor.

• Type II Error in statistical hypothesis testing
• Calculated as: FAR = (False Acceptances / Total Impostor Attempts)
• Lower FAR indicates stronger security posture
• Directly trades off against FRR; tightening the decision threshold to reduce FAR inevitably increases FRR

The probability that a biometric system incorrectly rejects a legitimate, authorized user or device. FRR measures system usability and convenience—a high FRR frustrates genuine users.

• Type I Error in statistical hypothesis testing
• Calculated as: FRR = (False Rejections / Total Genuine Attempts)
• High FRR leads to user abandonment and operational friction
• Loosening the threshold to reduce FRR increases FAR, creating the fundamental security-convenience trade-off

A graphical plot that visualizes the trade-off between FAR and FRR across all possible decision thresholds. The DET curve is the standard tool for evaluating biometric system performance.

• X-axis typically plots FAR, Y-axis plots FRR on a normal deviate scale
• The curve sweeps from strict (low FAR, high FRR) to lenient (high FAR, low FRR) thresholds
• EER is the point where the DET curve intersects the diagonal line where FAR equals FRR
• A curve closer to the origin indicates superior overall system accuracy

A related but distinct performance visualization that plots True Acceptance Rate (1-FRR) against FAR. Unlike the DET curve, ROC curves show performance from the perspective of genuine user success.

• Y-axis: True Acceptance Rate (TAR) or Genuine Acceptance Rate (GAR)
• X-axis: False Acceptance Rate (FAR)
• Area Under the Curve (AUC) provides a single aggregate performance metric
• Often preferred in machine learning contexts over DET curves for classifier evaluation

The process of selecting the similarity score cutoff that determines whether a biometric sample is accepted as a match. This threshold directly controls the FAR-FRR balance.

• A high threshold demands near-perfect matches, minimizing FAR but increasing FRR
• A low threshold accepts more variation, reducing FRR but increasing FAR
• Application-specific tuning is essential: banking apps prioritize low FAR, while convenience apps prioritize low FRR
• EER provides a threshold-independent metric for comparing systems before application-specific tuning

The alignment of a classifier's output probability with the true likelihood of correctness. Well-calibrated scores are essential for meaningful FAR and FRR calculations.

• A score of 0.95 should mean the system is correct 95% of the time
• Poor calibration distorts the DET curve and renders EER misleading
• Techniques like Platt scaling and isotonic regression correct miscalibration
• Critical for few-shot device enrollment where limited data can produce overconfident predictions