In Jeff Richter's book CLR via C# on page 396 is the best description of attributes I have read yet:

I've found this to be the best way for developers to think of custom attributes: instances of classes that have been serialized to a byte stream that resides in metadata.  Later, at run time, an instance of the class can be constructed by deserializing the bytes contained in the metadata.  In reality, what actually happens is that the compiler emits the information necessary to create an instance of the attribute class into metadata.  Each constructor parameter is written out with a 1-byte type ID followed by the value.  After “serializing“ the constructor's parameters, the compiler emits each of the specified field and property values by writing out the field/property name followed by a 1-byte type ID and then the value.  For arrays, the count of elements is saved first, followed by each individual element.

If you don't have this book you really should go buy it and read it from cover to cover.