Domain generalization approaches aim to learn a domain invariant prediction\nmodel for unknown target domains from multiple training source domains with\ndifferent distributions. Significant efforts have recently been committed to\nbroad domain generalization, which is a challenging and topical problem in\nmachine learning and computer vision communities. Most previous domain\ngeneralization approaches assume that the conditional distribution across the\ndomains remain the same across the source domains and learn a domain invariant\nmodel by minimizing the marginal distributions. However, the assumption of a\nstable conditional distribution of the training source domains does not really\nhold in practice. The hyperplane learned from the source domains will easily\nmisclassify samples scattered at the boundary of clusters or far from their\ncorresponding class centres. To address the above two drawbacks, we propose a\ndiscriminative domain-invariant adversarial network (DDIAN) for domain\ngeneralization. The discriminativeness of the features are guaranteed through a\ndiscriminative feature module and domain-invariant features are guaranteed\nthrough the global domain and local sub-domain alignment modules. Extensive\nexperiments on several benchmarks show that DDIAN achieves better prediction on\nunseen target data during training compared to state-of-the-art domain\ngeneralization approaches.\n