The　proliferation　of　Internet　of　Things　(IoT)　devices　in　the　5G　era　has　resulted　in　increased　security　vulnerabilities　and　zero-day　attacks,　underscoring　the　importance　of　network　intrusion　detection　systems　(NIDS).　However,　existing　NIDS　have　limitations　in　terms　of　accuracy,　recall　rates,　false　alarm　rates,　and　generalization　capabilities,　and　they　cannot　meet　the　IoT's　requirements　for　low　latency　and　limited　computing　resources.　To　overcome　these　challenges,　we　propose　a　NIDS　based　on　a　pseudo-siamese　stacked　autoencoder　(PSSAE),　deployed　in　the　fog　computing　layer.　Our　system　uses　unsupervised　training　of　stacked　autoencoders　(SAEs)　to　extract　deep　semantic　features　of　normal　and　abnormal　traffic,　followed　by　supervised　learning　with　labels　to　improve　characterization　and　classification　capabilities.　The　results　show　that　our　proposed　method's　accuracy　and　detection　rate　(DR)　is　2%　to　15%　and　1%-14%　higher　than　the　existing　techniques　using　the　KDDTest+　dataset,　respectively.　Our　proposed　method　outperformed　the　existing　methods　by　1%　to　4%　using　the　KDDTest+　dataset.　The　F1-Score　is　higher　by　3%　-　11.55%　using　the　KDDTest+　dataset.　On　the　other　hand,　using　the　KDDTest-21　dataset,　the　accuracy　of　our　proposed　method　also　outperformed　the　existing　technique　by　6.09%　-　13.81%.　The　DR　and　F1-Score　are　higher　by　7.02%　and　5.57%,　respectively,　using　the　KDDTest+　dataset.　This　is　due　to　the　fact　that　each　layer　of　the　network　trained　by　SAEs　is　more　capable　of　extracting　the　semantic　features　of　the　data　than　the　DNN-trained　network　directly.　IEEE