The integration of deep learning models into healthcare, finance, and government sectors has significantly advanced the capabilities of artificial intelligence (AI). However, these deep learning models require substantial computational power, leading many organizations to rely on cloud-based services for model training. This shift has raised critical privacy concerns because sensitive data used in training may be exposed to privacy attacks. This survey systematically examines the state of privacy attacks against deep learning and the effectiveness of current privacy-preserving methodologies. Using a PRISMA-style review, we analyze papers from the recent five-year literature and present (1) a unified taxonomy of attack families (membership inference, model inversion, model extraction, gradient leakage, property/attribute inference, poisoning/backdoors, side-channels, and LLM-specific leakage); (2) a taxonomy of defenses spanning differential privacy, federated learning (and robust aggregation), cryptographic protocols (HE/SMPC), anonymization & sanitization, PATE/shuffle models, and hybrid stacks; (3) recommended evaluation metrics for each attack family; and (4) a reproducibility checklist and an attack×defense protection matrix that synthesizes the literature’s qualitative findings. We complement these contributions with three case studies analysis of current landscape of privacy-preserving defenses against privacy attacks in centralized image classification, federated learning, and LLMs, which highlight practical trade-offs between privacy, utility, and system cost. Our analysis finds that noise-based methods such as DP remain the practical baseline but offer only partial protection; cryptographic approaches provide stronger theoretical guarantees at substantially higher cost; and LLM/multimodal leakage remains an urgent, under-benchmarked gap.