EaZy Learning: Adaptive Ensemble for Fingerprint Liveness

• EaZy Learning is an adaptive ensemble technique that leverages disjoint clusters and weighted majority voting to dynamically adjust to dataset heterogeneity.
• It partitions data via clustering to develop specialized base classifiers, effectively bridging the gap between global eager and local lazy learning strategies.
• Empirical evaluations demonstrate that EaZy Learning outperforms conventional methods in cross-sensor and cross-dataset fingerprint liveness detection.

EaZy Learning refers to an adaptive ensemble learning paradigm introduced for robust fingerprint liveness detection, primarily to improve generalization under cross-sensor and cross-dataset conditions—a stance where traditional and even standard ensemble-based classifiers historically exhibit significant limitations. The methodology is founded on partitioning the dataset according to its intrinsic structure, constructing an ensemble of disjoint, cluster-specific base classifiers, and integrating their outputs via performance-weighted voting. This generative process bridges the gap between eager and lazy learning paradigms, yielding an adaptable learning architecture that tunes its hypothesis granularity to the heterogeneity of the input data (Agarwal et al., 2021).

1. Motivation: Adaptive Learning Between Eager and Lazy Paradigms

Conventional eager learning methods, such as SVMs and neural networks, train a single global predictor over the entire dataset, resulting in a model that is typically fast at test time but insufficiently adaptive to local feature variation. Lazy learning approaches like k-NN, conversely, perform prediction by instance retrieval and aggregation, facilitating superior local adaptation at the expense of increased query-time computational complexity and storage.

Standard ensemble learning strategies (e.g., bagging, boosting) are often static in both ensemble size and the scope of base learners, thus unable to align learning granularity with dataset compositional complexity. EaZy Learning circumvents these issues by:

• Exploiting local structure via data-driven clustering and local hypothesis construction.
• Compactly representing each cluster by a distinct base classifier, removing the necessity to retain all training points.
• Automatically scaling the number of base classifiers based on measured dataset heterogeneity; homogeneous datasets reduce to a single classifier (eager regime), while highly diverse datasets result in numerous, specialized classifiers (approximating lazy behavior).

2. Construction of the EaZy Learning Ensemble

EaZy Learning formalizes ensemble construction through disjoint cluster partitions:

• An ensemble $\Pi = \{\psi_1,\dots,\psi_n\}$ is generated, with each base classifier $\psi_i$ trained on a unique data subset $D_{\psi_i} \subset D$ where $D_{\psi_i} \cap D_{\psi_j} = \emptyset$ for all $i \neq j$ and $\bigcup_{i=1}^n D_{\psi_i} = D$.
• Each base classifier $\psi_i$ instantiates a mapping $\psi : H \times T \rightarrow Y$ for hypothesis space $H$, input space $T$, and label set $Y$ (e.g., $Y$ = {live, spoof}).

Clusters $\{c_1, ..., c_n\}$ are determined by a chosen clustering algorithm (e.g., Expectation Maximization), and within each, a base learner $K$ (e.g., Sequential Minimal Optimization, SMO) is fit to the data.

3. Weighted Majority Voting and Validation

Integration of predictions is performed via a weighted voting mechanism, with weights determined from hold-out validation accuracy:

$$W_i = \frac{A_i}{\sum_{j=1}^n A_j}, \quad \text{where} \quad A_i = \frac{|\{x \in V : \psi_i(x) = \text{true label}\}|}{|V|}$$

In prediction, for query $x_q$, the system assigns the label maximizing the total weight of supporting classifiers:

$$\hat{y} = \underset{y \in \{\text{live, spoof}\}}{\arg\max} \sum_{i: \psi_i(x_q) = y} W_i$$

4. Training and Prediction Algorithms

Algorithm 1 (Training):

1. Randomly partition the full training set $D$ into $D_{train}$ and hold-out set $V$.
2. Cluster $D_{train}$ into $\{c_1, ..., c_n\}$.
3. For each cluster $c_i$, train $\psi_i = K(c_i)$.
4. Evaluate each $\psi_i$ on $V$, compute accuracy $A_i$.
5. Set weights $W_i = A_i / (\sum_j A_j)$.

Algorithm 2 (Prediction):

1. For a query $x_q$, compute $s(y) = \sum_{i: \psi_i(x_q) = y} W_i$ for each label $y$.
2. Return $\hat{y} = \arg\max_y s(y)$.

5. Computational Complexity and Adaptivity

The clustering step scales with the clustering algorithm (for EM: $O(In d^2)$ where $I$ = iterations, $n$ = instances, $d$ = features). Base model training is typically no more demanding than a single global model, and test-time cost is $O(n \cdot \text{cost}(\psi_i(x_q)))$. Notably, $n$ is determined by data structure, enabling natural scaling; homogeneous datasets yield $n \approx 1$, whereas heterogeneous data lead to larger $n$. This adaptivity balances inferential efficiency with modeling expressiveness.

6. Empirical Evaluation and Comparative Performance

Experiments were conducted on LivDet 2011, 2013, and 2015 using features derived from ResNet-50 embeddings (2048-dimensional). Two primary evaluation regimes were used:

• Cross-sensor: Training and testing on different sensors within a given LivDet year.
• Cross-dataset: Training on one year’s dataset, testing on another.

Performance was assessed using overall accuracy and APCER (Attack Presentation Classification Error Rate). Results, averaged across multiple splits, are summarized as:

Method	Cross-Sensor Acc. (%)	Cross-Sensor APCER	Cross-Dataset Acc. (%)	Cross-Dataset APCER
EaZy Learning	65.89	0.44	60.49	0.18
Random Subspace (SMO)	65.76	0.31	55.70	0.74
Bagging (SMO)	64.83	0.27	55.49	0.74
AdaBoost	64.12	0.30	56.43	0.68
Random Forest	61.88	0.47	50.94	0.79

Statistically, a Friedman test yields $\chi^2_r=18.225$, $p=0.00112 < 0.05$, indicating EaZy Learning’s cross-dataset APCER is significantly superior to competitors (Agarwal et al., 2021).

7. Diversity, Generalization, and Conclusions

Disjoint clustering ensures each $\psi_i$ specializes in a low-correlation region of feature space, generating high ensemble diversity. Weighted voting prioritizes robust, generalizable hypotheses. A direct consequence is an ensemble size $n$ that interpolates between global and local error bounds, depending on the data’s complexity. In practice, this manifests as superior open-set generalization—significantly better cross-sensor and cross-dataset performance than classical bagging, boosting, or random forests. A plausible implication is that such model adaptivity and specialization strategies could extend to generalizable ensemble techniques for other biometric or heterogeneous data analysis domains (Agarwal et al., 2021).