It　is　widely　acknowledged　that　Deep　Neural　Networks　(DNNs)　are　vulnerable　to　backdoor　attacks,　wherein　attackers　inject　poisoned　samples　into　the　training　data,　resulting　in　misclassification　of　these　tainted　inputs　during　model　deployment.　Given　that　DNNs　typically　utilize　external　training　data　from　untrusted　third　parties,　it　becomes　imperative　to　employ　a　robust　defense　strategy　against　backdoor　attacks　during　the　training　phase.　In　this　paper,　we　propose　a　novel　framework　for　Dual-Model　Backdoor　Defense　(DMBD)　for　defending　against　backdoor　attacks　,　which　enables　learning　of　resilient　representations　and　processing　of　poisoned　samples　for　accurate　classification.　Our　defense　framework　comprises　two　steps:　Firstly,　disrupting　the　established　relationship　between　original　triggers　and　target　labels　by　introducing　additional　triggers　to　all　images　and　applying　multiple　data　enhancements;　Secondly,　employing　a　contrastive　learning　model　to　associate　unlabeled　latent　variables　in　the　Gaussian　Mixture　Model　(GMM)　with　suspicious　label　annotations.　By　considering　the　underlying　data　distribution,　mislabeled　samples　are　identified　as　out-of-distribution　instances　using　a　two-component　GMM　and　subsequently　subjected　to　cross-supervised　loss　for　re-labeling　and　further　processing.　Extensive　experiments　demonstrate　that　our　proposed　backdoor　defense　(DMBD)　consistently　delivers　stable　performance　across　various　attack　scenarios.　©　2025　IEEE.