Security of watermark is a huge concern when dealing with imperceptible watermarking for copyright protection and owner identification. A novel watermarking technique that is intrinsically secure against unauthorized extraction of watermark and robust against a variety of attacks has been proposed in this paper. Semi-blind in nature, the technique employs fractional discrete cosine transform (FDCT) and zig-zag scrambling to add a layer of security while embedding watermark bits in the cover image. Transformation through multiple fractional value pairs yields robust results against attacks for every cover image. A pair out of these is chosen while embedding the watermark. The same pair is needed at the time of extraction of original watermark from a signed image. This pair of unique values adds a layer of security in the proposed technique. After dividing the cover image into sub-blocks that are non-overlapping, fuzzy entropy is used to make the selection of appropriate sub-blocks in which watermark bits are to be embedded. These blocks are first mapped in a zig-zag fashion and then, after selection of half sub-blocks, FDCT is applied on each of them. FDCT coefficients serve as training data for Kernel extreme learning machine (KELM). Used as a non-linear regression model, KELM predicts the value of a particular coefficient with the help of neighboring transform domain coefficients. Multiple scaling factors are used to control the extent to which the watermark bit modifies the predicted coefficient. Strength of these scaling factors is determined and controlled with the help of standard deviation of every selected sub-block. The technique shows perfect extraction of watermarks (Normalized correlation=1.0, Bit error ratio=0.0) against attacks like median filtering (3 × 3), Gaussian filtering, Weiner filtering, JPEG compression (Quality factor 40) and gamma correction. Even for other attacks, the extracted watermarks are not too bad in quality that they are beyond recognition.