Evaluation of Variational Quantum Classifiers (VQC) for Cyberattack Detection in the NISQ Era

We investigate the structural limits of Variational Quantum Classifiers (VQC) for network anomaly detection in the Noisy Intermediate-Scale Quantum (NISQ) era. Using the official 20% research subset of NSL-KDD, we train a 4-qubit classifier that embeds 16 principal components by amplitude encoding and reaches 88% accuracy on the held-out partition. On the four established attack families, an explicit many-to-one decoding of the sixteen basis outcomes attains 71.8% in-sample accuracy and a macro-averaged recall of 0.787 on held-out attack records, against 0.250 for a majority-class classifier. Aggregate accuracy conceals that separation entirely: 78.4% for the trained model against 78.7% for the trivial one. Measuring class separability in the encoded states before training, amplitude encoding retains an AUC of 0.560±0.018 at sixteen components over five seeds, against 0.762±0.020 for angle encoding. Yet a trained comparison at matched parameter count on four components converts none of that advantage into accuracy: the separation is present in the encoded geometry and absent from the classifier, which implicates the readout alongside the encoding, a distinction this design does not separate. A controlled sweep over five ansatz depths and five seeds, spanning 8 to 72 trainable parameters at fixed encoding, data and decoding, finds subsample accuracy saturating at 0.880±0.005 while the training objective keeps falling, with a parameter-matched circuit trailing a 73-parameter classical network by 7.9 points across the two protocols compared. Depth is therefore not the binding constraint. On KDDTest+, classical recall on novel signatures falls by 19 to 32 points while the variational classifier shows no comparable decline.