You really want to encrypt something, translate into one language, and retranslate the syllables into a phonetic character set.
Would it even be able to translate that back?
You'd have to understand the Kanji's pronunciation to hear what they are, then translate that spoken phrase into the original language. If directly translated, you get hilarious Engrish, or total and utter gibberish. Mix-and-match other languages all over the place, and tier relative phonetics, and you've got mayhem to translate. Cross-translations go even more hectic, adding even more encryption per-language tier.
Fireball = Bola de fuego (Bo•la de fu•e•go or Bo la de fu e go) = ボーFUと行く (translated back to English is funny. (I'll go with Bo FU))
I tried to find a way to reverse the process; but it's easier said than done. Needs some perfection, but it's a good start. There doesn't happen to be a translator that converts roman characters to phonetic characters and back, is there (no = の, for example)? Because this isn't a proper encryption method to apply. With what little I already know of Japanese, it's already a broken translation. I recognize some characters that shouldn't be in that mix.
EDIT:
Here we go: (kinda a mixed-bag here (hirigana, kanji, and katakana). so a messy re-translation)
ぼ(la) で ふエご (DirTrans: "The pot (la) in Fuego")
ぼ (la) で ふ エ ご (can't find a compatible character for LA)
bo la de fu e go
bola de fuego
fireball
(might need extended character sets to see characters)
An alternate approach:
Dig Deeper
cavar más profundo
ca var mas pro fun do
kya va ma pyo fu do
きゃ ヴァ ま ピョ ふ ど
きゃヴァ ま ピョふど (structured like "cavar más profundo"; direct translation: "If Ke Cordova or Pyofu etc.")
Meh. It was worth a shot. If anything it's more linguistic context/syntax awareness that's necessary to make this work properly. Would make for a good riddle mechanic nonetheless. Like what if in an alternate universe, this was how the Spanish language was written, instead of Roman/Latin characters?