Given the significant growth of cyberattacks, it becomes urgent to develop effective methods for identifying anomalies in computer networks. Traditional security mechanisms, which predominantly rely on static signatures, often struggle to detect novel or sophisticated threats such as zero-day attacks and polymorphic malware. Consequently, the research community is increasingly turning towards data-driven approaches capable of learning dynamic traffic patterns to distinguish between benign behavior and malicious intent with high precision. This work benchmarks four supervised machine learning algorithms: SVM, Random Forest, LSTM, and Logistic Regression. These were applied to the recognition of malicious traffic using the CICIDS2017 dataset, which is known for its diverse set of attacks. The methodology adopted includes preprocessing, normalization, feature selection, and hyperparameter tuning for each algorithm to ensure greater reliability in the results. The F1- score metric was used for model evaluation due to the original data imbalance on the CICIDS2017 data. This methodological choice aims to minimize common distortions in real-world environments, where attacks typically constitute a minority of traffic compared to legitimate traffic. The experimental results indicated that all the algorithms obtained high F1-scores in the {0.937; 0.999} interval, with the best results obtained by the RF-induced model. These findings reinforce the potential of combining machine learning techniques with updated datasets to create robust intrusion detection systems, contributing to the security of modern networks.