I'm a bit confused about just what you want to delete and what you want to keep, but here's an approach. I've no idea about its speed. It needs Perl version 5.10+. Note that field offsets are zero-based.

c:\@Work\Perl\monks>perl -wMstrict -le
"use 5.010;
 ;;
 my $inrec = '0;1;2?foo=DUH;3;4;5;6?bar=BLAB DAB;7;8;9;;;;;;';
 print qq{begin: '$inrec'};
 ;;
 my $sep   = ';';
 my $data  = qr{ [^\Q$sep\E] }xms;
 my $field = qr{ $data* \Q$sep\E }xms;
 ;;
 my %targets = qw(foo 2  bar 6);
 my @words =
   sort { $targets{$a} <=> $targets{$b} }
   keys %targets
   ;
 ;;
 for my $word (@words) {
   $inrec =~ s{ \A (?:$field){$targets{$word}}
                [^?]+ \? \Q$word\E = \K ($data*)
              }
              {}xms;
   $inrec .= qq{$1$sep};
   print qq{'$inrec'};
   }
 ;;
 print qq{final: '$inrec'};
"
begin: '0;1;2?foo=DUH;3;4;5;6?bar=BLAB DAB;7;8;9;;;;;;'
'0;1;2?foo=;3;4;5;6?bar=BLAB DAB;7;8;9;;;;;;DUH;'
'0;1;2?foo=;3;4;5;6?bar=;7;8;9;;;;;;DUH;BLAB DAB;'
final: '0;1;2?foo=;3;4;5;6?bar=;7;8;9;;;;;;DUH;BLAB DAB;'
[download]

Update: Changed code example to make more consistent use of the  $sep scalar.