毕业生知识库

随着移动互联网的高速发展，人类社会快速信息化。作为信息社会的基础设施，身份识别技术通常是安全防护系统的核心环节。生物特征识别技术摆脱了对特定标志物和特定知识的依赖，更具普遍性、唯一性、安全性和高效性，是新型身份识别技术的代表。现有的生物特征识别系统通常应用于受控环境，即样本采集的大部分因素都受到人为控制，例如采集距离、采集光照、被识别对象的姿态、是否佩戴配饰等。然而随着经济社会的发展，能够适应非受控环境的、更为安全可靠和高效便捷的身份识别方法成为了现代社会发展的迫切需求。在非受控环境中，上述样本采集过程中的诸多因素将不再受到人为控制，或受控的程度 大为减少，导致采集得到的样本出现诸多变化，给生物特征识别系统带来了一系列新的重大挑战，主要包括:(1)用户非配合条件下的难样本，(2)对抗攻击带来的安全威胁，(3)复杂场景中遮挡模糊等因素导致的样本难以感知。因此，针对这些挑战开展面向非受控环境的新型生物特征识别技术的研究具有重要的应 用价值。同时也将推动模式识别、计算机视觉等前沿领域的进一步发展，具有较 强的理论和学术意义。

论文在以非受控环境中的应用需求为牵引，采用图模型作为基本技术路线，针对生物特征识别的三个核心模块:样本预处理、特征表达、特征匹配提出了新的解决方案。首先针对对抗攻击带来的安全威胁，提出了基于可学习的图嵌入的样本预处理方法，抑制对抗噪声带来的不良影响;采用动态图表达改变生物特征表达的原有范式，以适应非配合难样本的识别与匹配;最后，采用概率图模型对复杂的场景信息进行挖掘和利用，有效缓解了复杂场景中可用信息不足的问题，在特征匹配环节实现场景信息与知识与视觉特征相似图的有效融合。

论文的主要工作和创新点可以归纳如下:

1. 基于可学习的图嵌入的对抗防御方法。面对非受控环境中恶意对抗攻击带来的严重安全威胁，论文首先分析了生物特征识别中对抗样本的特点。对抗攻击方法通过在样本中添加对抗噪声，破坏了样本空间的局部邻接关系。基于此，论文采用图方法来建模样本空间的结构，尤其是样本局部邻域的结构。通过将样本建模为图的顶点，样本之间的局部邻接关系建模为图的边，论文提出了一种基于可学习的图嵌入的对抗防御方法。其核心思路是通过寻找一个恰当的图嵌入空间，恢复样本之间的局部邻接关系，从而有效抑制对抗样本带来的不良影响，提升对抗攻击的难度和成本。在对图嵌入空间进行估计时，为了解决局部嵌入空间选择不可微的问题，论文采用强化学习的方式进行训练，实现了可学习的图嵌入方法。论文将图嵌入方法与深度学习方法进行了有机融合，对其进行了继承和发扬，并在多种实验条件下的实验结果验证了这一方法的有效性。

2. 基于动态图模型的生物特征表达范式。在非受控环境中部署的生物特征识别系统，往往难以要求识别对象的进行高度配合。这就使得采集得到的图像容易受到诸多干扰因素(例如光照、姿态、遮挡等)的不利影响，导致一般的深度学习模型的特征表达能力大为下降，其中以遮挡最为严重。为此，论文提出了一种新颖的特征表达方法，将生物特征样本表达为动态图。图的每个顶点对应于样本的特定子区域，而边则表示子区域之间的结构关系。在匹配过程中，论文通过对图表示进行动态调整实现更好的特征表达能力。这一方法并不局限于深度学习的数据驱动的模式，论文将其与手工设计的特征提取方法进行了试验性融合，使其成为了一种与具体的特征提取方法无关的特征表达范式。在虹膜和人脸两种生物特征模态上进行的较为完备的实验证明了这一范式能够大大提升非配合条件下的特征表达能力。

3. 基于概率图模型融合场景感知的特征匹配方法。非受控环境下的应用场景往往较为复杂，场景本身对识别系统的影响也大大加强。尤其是在行人在识别任务中，在大规模场景下，样本本身的信息量往往已经不足以支撑应用需求。为此，论文提出采用概率图模型对场景信息进行描述和利用，同时将实例级状态信息引入模型，提供更为精细的描述与预测。最终通过自适应的联合度量，论文将场景信息和视觉信息进行了有效地融合，缓解了复杂场景中样本信息量不足的问题，使得生物特征识别系统的精度和效率同时得到有效提升。

Identification technology is the infrastructure of the information society. Biometric technology, which is free from specific markers and knowledge, is the most promising candidate. Most of the existing biometric recognition systems are deployed in a con- trolled environment, in which the acquisition distance, illumination, the pose of the object, and whether wearing accessories or not are controlled. With the development of the mobile Internet and the informatization of human society, there is an urgent need for more secure, reliable, efficient, and convenient identification methods in uncontrolled environments. However, existing biometric methods are not yet able to address chal- lenges posed by uncontrolled environments: (1) hard samples under non-cooperative conditions, (2) security threats posed by adversarial attacks, and (3) information insuf- ficiency due to complex scenarios. Therefore, research on novel biometric methods for uncontrolled environments to address these challenges is of great application value. It will also promote the development of pattern recognition and computer vision.

To meet the challenges in uncontrolled environments, this dissertation adopts the graph model as the foundation, since the characteristics of the graph model dovetail with the technical requirements. Firstly, a learnable graph embedding method is pro- posed to suppress the adverse effects of adversarial noises during pre-processing. Then, the original paradigm of feature representation is replaced by dynamic graph represen- tation to improve the performance of hard samples under non-cooperative conditions. Finally, the probabilistic graph model is used to mine the scene information during fea- ture matching, which alleviates the information insufficiency in complex scenes.

The major contributions of this dissertation can be summarized as follows:

1. Learnable graph embedding based adversarial defense. This dissertation first analyzes the characteristics of the adversarial samples against biometric recogni- tion systems in uncontrolled environments. Adversarial noises break the local adja- cency between samples. Hence, this dissertation adopts the graph method to model the local adjacency between samples. The samples are represented by nodes, the adjacent relationship is represented by the adjacent matrix. In order to recover the adjacent re- lationship of samples, this dissertation proposes a learnable graph embedding method, which estimates a graph embedding space for samples. The adverse effects of adver- sarial noises are restrained in the graph embedding space. The proposed method adopts deep reinforcement learning to estimate the proper graph embedding space. The graph embedding method is integrated with the deep learning methods. The experimental results under a variety of experimental conditions validate the effectiveness of the pro- posed method.

2. Dynamic graph based feature representation. Most of the identified subjects are non-cooperative in uncontrolled environments. Therefore, the acquired samples are more likely to be degraded by disturbing factors, such as illumination, posture, and espe- cially occlusions. To this end, this dissertation proposes a novel feature representation method, which adopts dynamic graphs as the representation of samples. Each vertex of the graph corresponds to a specific subregion of the sample, while the edges repre- sent the structural relationships between the subregions. The graph representation is dynamically adjusted during matching. This approach is not limited to the data-driven paradigm of deep learning, making it a novel feature representation paradigm that is independent of the feature extraction method. Complete experiments on two modalities demonstrate its effectiveness under non-cooperative conditions.

3. Probabilistic graphical model based feature matching. The impact of appli- cation scenarios in uncontrolled environments on the recognition system is significant. The amount of information in the sample itself is no longer sufficient to support the application requirements in some cases. To this end, this dissertation proposes to adopt the probabilistic graphical model for the description and utilization of scene informa- tion. Instance-level state information is introduced into the model to provide a more fine-grained description. And the scene information and visual information are fused by an adaptive joint metric during feature matching. In this way, the problem of infor- mation insufficiency in complex scenes is alleviated, and the accuracy and efficiency of the recognition system are effectively improved at the same time.