Donald Broadbent's 1958 book Perception and Communication is the oldest work I can find that makes use of this term.
However, I wonder if perhaps someone else had used the term earlier, or if Broadbent had used the term in published work before that date.
Colin Cherry coined the phrase the Cocktail party effect to describe the ability for people to tune or filter out distraction and focus on one thing; the illustration being able to focus on a conversation amidst all the background noise at a cocktail party.
Broadbent's filter model of attention was developed from Cherry's work proposing the concept filtering information. Cherry first used the idea of a filter. Broadbent was the first person to officially coin the term filter as it is used within this context and within a cognitive model.
as others have mentioned, Colin Cherry's cocktail party problem highlighted a crucial bottleneck in information flow at the highest levels of sensory-motor neural integration. by giving this phenomenon a sexy title Cherry showed an intuitive gift for public relations! however the ground for his thinking was prepared by prior theoretical developments - from the point of view of psychology one should mention Kenneth Craik's work during WW2. undoubtedly the seminal event was the publication at the end of the 1940's of Shannon and Weaver's Mathematical Theory of Communication. this book codified a set of ideas which had been arising naturally in many different areas of study due to the rapidly increasing impact of automatic computation, control systems and electronic communications.
as a footnote, it may merit attention that Donald Broadbent's early use of the term filter in 1958 occurred at pretty much the same epoch as the mathematically sophisticated work of Rudolf Kalman (cf Kalman Filter) and co-workers was finding its way into the literature.
the rather close semantic association between the new "sciences" of cybernetics and ergonomics, as also the intimate connection of both with what engineers were coming to know as automatic control theory should alert us to the fact that it would be naive to view this paradigm within psychology as a form of pure research. in fact it has always been very task-oriented, and relates to the increasing role of human beings as components of complex systems in the context of large-scale engineering, industrial and military applications. Kalman's team was funded by the US Navy, whilst the MRC Psychology Unit at Cambridge was always a government research department run on civil service lines rather than an academic institution.
this background may explain the otherwise curious fact that in the 1950's a huge research effort was directed towards questions relating to vigilance. interesting as it may be, the experimental vigilance paradigm would seem an unlikely choice of topic if we wished to interpret human information processing research in terms of the task of developing a unifying model of mammalian behaviour.
an event-filled decade or so of detailed empirical research and conceptual development in the sphere of human information processing is neatly epitomized by looking at Broadbent's groundbreaking Perception and Communication (1958) and comparing it with his comprehensive review of the field in Decision and Stress (1971)
