Abstract: In this study, a novel artificial immune system algorithm for unsupervised classification and recognition is proposed by using a novel manifold distance based dissimilarity measure which can measure the geodesic distance along the manifold. The new method formulizes the immune response as a quaternion AIR=(G,I,R,A), where G denotes exterior stimulus or antigen, I denotes the set of valid antibodies, R denotes the set of reaction rules describing the interactions between antibodies, and A denotes the dynamical algorithm describing how the reaction rules are applied to antibody population. In order to solve unsupervised classification problems, the new method encodes each antibody as a sequence of real integer numbers representing the cluster representatives, and searches the optimal cluster representatives from a combinatorial optimization viewpoint using the dynamical algorithm A. Experimental results on six artificial datasets with different manifold structures and the USPS handwritten digit datasets show that the novel algorithm has the ability to identify complex non-convex clusters, compared with the K-means algorithm, a genetic algorithm-based clustering proposed by Maulik, and an evolutionary clustering algorithm with the manifold distance.