Many audio applications perform perception-based time-frequency (TF) analysis by decomposing sounds into a set of functions with good TF localization (i.e. with a small essential support in the TF domain) using TF transforms and applying psychoacoustic models of auditory masking to the transform coefficients. To accurately predict masking interactions between coefficients, the TF properties of the model should match those of the transform. This involves having masking data for stimuli with good TF localization. However, little is known about TF masking for mathematically well-localized signals. Most existing masking studies used stimuli that are broad in time and/or frequency and few studies involved TF conditions. Consequently, the present study had two goals. The first was to collect TF masking data for well-localized stimuli in humans. Masker and target were 10-ms Gaussian-shaped sinusoids with a bandwidth of approximately one critical band. The overall pattern of results is qualitatively similar to existing data for long maskers. To facilitate implementation in audio processing algorithms, a dataset provides the measured TF masking function. The second goal was to assess the potential effect of auditory efferents on TF masking using a modeling approach. The temporal window model of masking was used to predict present and existing data in two configurations: (1) with standard model parameters (i.e. without efferents), (2) with cochlear gain reduction to simulate the activation of efferents. The ability of the model to predict the present data was quite good with the standard configuration but highly degraded with gain reduction. Conversely, the ability of the model to predict existing data for long maskers was better with than without gain reduction. Overall, the model predictions suggest that TF masking can be affected by efferent (or other) effects that reduce cochlear gain. Such effects were avoided in the experiment of this study by using maximally-compact stimuli.